From 5e04ec70a3d0e96e529343070b98d3f4d2936643 Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Mon, 26 Nov 2012 11:52:11 +0800 Subject: [PATCH 01/83] make a 'real' fake data --- config/mongoid.yml | 2 +- db/data | 2429 ++++++++++++++++++++++++++++++++++++++++++++ db/seeds.rb | 72 +- 3 files changed, 2480 insertions(+), 23 deletions(-) create mode 100644 db/data diff --git a/config/mongoid.yml b/config/mongoid.yml index 75c5ef80..a809bc06 100644 --- a/config/mongoid.yml +++ b/config/mongoid.yml @@ -2,7 +2,7 @@ defaults: &defaults host: localhost # slaves: # - host: slave1.local - port: 37017 + port: 27017 # - host: slave2.local # port: 27019 diff --git a/db/data b/db/data new file mode 100644 index 00000000..98407fac --- /dev/null +++ b/db/data @@ -0,0 +1,2429 @@ +[ + { + "author": [ + "Liu, Wei", + "Wang, Jun", + "Kumar, Sanjiv", + "Chang, Shih-Fu" + ], + "paper_title": "Hashing with Graphs ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1, + "page_to": 8, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Hashing is becoming increasingly popular for efficient nearest neighbor search in massive databases. However, learning short codes that yield good search perfromance is still a challenge. Moreover, in many cases real-world data lives on a low-dimensional manifold, which should be taken into account to capture meaningful nearest neighbors. In this paper, we propose a novel graph-based hashing method which automatically discovers the neighborhood structure inherent in the data to learn appropriate compact codes. To make such an approach computationally feasible, we utilize Anchor Graphs to obtain tractable low-rank adjacency matrices. Our fromulation allows constant time hashing of a new data point by extrapolating graph Laplacian eigenvectors to eigenfunctions. Finally, we describe a hierarchical threshold learning procedure in which each eigenfunction yields multiple bits, leading to higher search accuracy. Experimental comparison with the other state-of-the-art methods on two large datasets demonstrates the efficacy of the proposed method.\n" + }, + { + "author": [ + "Zhong, Wenliang", + "Kwok, James" + ], + "paper_title": "Efficient Sparse Modeling with Automatic Feature Grouping", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 9, + "page_to": 16, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The grouping of features is highly beneficial in learning with high-dimensional data. It reduces the variance in the estimation and improves the stability of feature selection, leading to improved generalization. Moreover, it can also help in data understanding and interpretation. OSCAR is a recent sparse modeling tool that achieves this by using a $\\ell_1$-regularizer and a pairwise $\\ell_\\infty$-regularizer. However, its optimization is computationally expensive. In this paper, we propose an efficient solver based on the accelerated gradient methods. We show that its key projection step can be solved by a simple iterative group merging algorithm. It is highly efficient and reduces the empirical time complexity from $O(d^3 \\sim d^5)$ for the existing solvers to just $O(d)$, where $d$ is the number of features. Experimental results on toy and real-world data sets demonstrate that OSCAR is a competitive sparse modeling approach with the added ability of automatic feature grouping.\n" + }, + { + "author": [ + "Bi, Wei", + "Kwok, James" + ], + "paper_title": "Multi-Label Classification on Tree- and DAG-Structured Hierarchies ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 17, + "page_to": 24, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Many real-world applications involve multi-label classification, in which the labels are organized in the from of a tree or directed acyclic graph (DAG). However, current research efforts typically ignore the label dependencies or can only exploit the dependencies in tree-structured hierarchies. In this paper, we present a novel hierarchical multi-label classification algorithm which can be used on both tree- and DAG-structured hierarchies. The key idea is to fromulate the search for the optimal consistent multi-label as the finding of the best subgraph in a tree/DAG. Using a simple greedy strategy, the proposed algorithm is computationally efficient, easy to implement, does not suffer from the problem of insufficient/skewed training data in classifier training, and can be readily used on large hierarchies. Theoretical results guarantee the optimality of the obtained solution. Experiments are perfromed on a large number of functional genomics data sets. The proposed method consistently outperfroms the state-of-the-art method on both tree- and DAG-structured hierarchies.\n" + }, + { + "author": [ + "He, Jingrui", + "Lawrence, Rick" + ], + "paper_title": "A Graph-based Framework for Multi-Task Multi-View Learning ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 25, + "page_to": 32, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Many real-world problems exhibit dual-heterogeneity. A single learning task might have features in multiple views (i.e., feature heterogeneity); multiple learning tasks might be related with each other through one or more shared views (i.e., task heterogeneity). Existing multi-task learning or multi-view learning algorithms only capture one type of heterogeneity. In this paper, we introduce Multi-Task Multi-View (M^2TV) learning for such complicated learning problems with both feature heterogeneity and task heterogeneity. We propose a graph-based framework (GraM^2) to take full advantage of the dual-heterogeneous nature. Our framework has a natural connection to Reproducing Kernel Hilbert Space (RKHS). Furthermore, we propose an iterative algorithm (IteM^2) for GraM^2 framework, and analyze its optimality, convergence and time complexity. Experimental results on various real data sets demonstrate its effectiveness.\n" + }, + { + "author": [ + "Zhou, Tianyi", + "Tao, Dacheng" + ], + "paper_title": "GoDec: Randomized Low-rank & Sparse Matrix Decomposition in Noisy Case ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 33, + "page_to": 40, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Low-rank and sparse structures have been profoundly studied in matrix completion and compressed sensing. In this paper, we develop ``Go Decomposition'' (GoDec) to efficiently and robustly estimate the low-rank part $L$ and the sparse part $S$ of a matrix $X=L+S+G$ with noise $G$. GoDec alternatively assigns the low-rank approximation of $X-S$ to $L$ and the sparse approximation of $X-L$ to $S$. The algorithm can be significantly accelerated by bilateral random projections (BRP). We also propose GoDec for matrix completion as an important variant. We prove that the objective value $\\|X-L-S\\|_F^2$ converges to a local minimum, while $L$ and $S$ linearly converge to local optimums. Theoretically, we analyze the influence of $L$, $S$ and $G$ to the asymptotic/convergence speeds in order to discover the robustness of GoDec. Empirical studies suggest the efficiency, robustness and effectiveness of GoDec comparing with representative matrix decomposition and completion tools, e.g., Robust PCA and OptSpace.\n" + }, + { + "author": [ + "Yu, Jia Yuan", + "Mannor, Shie" + ], + "paper_title": "Unimodal Bandits", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 41, + "page_to": 48, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We consider multiarmed bandit problems where the expected reward is unimodal over partially ordered arms. In particular, the arms may belong to a continuous interval or correspond to vertices in a graph, where the graph structure represents similarity in rewards. The unimodality assumption has an important advantage: we can determine if a given arm is optimal by sampling the possible directions around it. This property allows us to quickly and efficiently find the optimal arm and detect abrupt changes in the reward distributions. For the case of bandits on graphs, we incur a regret proportional to the maximal degree and the diameter of the graph, instead of the total number of vertices.\n" + }, + { + "author": [ + "Dinuzzo, Francesco", + "Ong, Cheng Soon", + "Gehler, Peter", + "Pillonetto, Gianluigi" + ], + "paper_title": "Learning Output Kernels with Block Coordinate Descent ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 49, + "page_to": 56, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose a method to learn simultaneously a vector-valued function and a kernel between its components. The obtained kernel can be used both to improve learning perfromances and to reveal structures in the output space which may be important in their own right. Our method is based on the solution of a suitable regularization problem over a reproducing kernel Hilbert space (RKHS) of vector-valued functions. Although the regularized risk functional is non-convex, we show that it is invex, implying that all local minimizers are global minimizers. We derive a block-wise coordinate descent method that efficiently exploits the structure of the objective functional. Then, we empirically demonstrate that the proposed method can improve classification accuracy. Finally, we provide a visual interpretation of the learned kernel matrix for some well known datasets.\n" + }, + { + "author": [ + "Minh, Ha Quang", + "Sindhwani, Vikas" + ], + "paper_title": "Vector-valued Manifold Regularization ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 57, + "page_to": 64, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We consider the general problem of learning an unknown functional dependency, f : X->Y, between a structured input space X and a structured output space Y, from labeled and unlabeled examples. We fromulate this problem in terms of data-dependent regularization in Vector-valued Reproducing Kernel Hilbert Spaces (Micchelli & Pontil, 2005) which elegantly extend familiar scalar-valued kernel methods to the general setting where Y has a Hilbert space structure. Our methods provide a natural extension of Manifold Regularization (Belkin et al., 2006) algorithms to also exploit output inter-dependencies while enforcing smoothness with respect to input data geometry. We propose a class of matrix-valued kernels which allow efficient implementations of our algorithms via the use of numerical solvers for Sylvester matrix equations. On multilabel image annotation and text classification problems, we find favorable empirical comparisons against several competing alternatives.\n" + }, + { + "author": [ + "Sugiyama, Masashi", + "Yamada, Makoto", + "Kimura, Manabu", + "Hachiya, Hirotaka" + ], + "paper_title": "On Infromation-Maximization Clustering: Tuning Parameter Selection and Analytic Solution ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 65, + "page_to": 72, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Infromation-maximization clustering learns a probabilistic classifier in an unsupervised manner so that mutual infromation between feature vectors and cluster assignments is maximized. A notable advantage of this approach is that it only involves continuous optimization of model parameters, which is substantially easier to solve than discrete optimization of cluster assignments. However, existing methods still involve non-convex optimization problems, and therefore finding a good local optimal solution is not straightforward in practice. In this paper, we propose an alternative infromation-maximization clustering method based on a squared-loss variant of mutual infromation. This novel approach gives a clustering solution analytically in a computationally efficient way via kernel eigenvalue decomposition. Furthermore, we provide a practical model selection procedure that allows us to objectively optimize tuning parameters included in the kernel function. Through experiments, we demonstrate the usefulness of the proposed approach.\n" + }, + { + "author": [ + "Nock, Richard", + "Magdalou, Brice", + "Briys, Eric", + "Nielsen, Frank" + ], + "paper_title": "On tracking portfolios with certainty equivalents on a generalization of Markowitz model: the Fool, the Wise and the Adaptive", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 73, + "page_to": 80, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Portfolio allocation theory has been heavily influenced by a major contribution of Harry Markowitz in the early fifties: the mean-variance approach. While there has been a continuous line of works in on-line learning portfolios over the past decades, very few works have really tried to cope with Markowitz model. A major drawback of the mean-variance approach is that it is approximation-free only when stock returns obey a Gaussian distribution, an assumption known not to hold in real data. In this paper, we first alleviate this assumption, and rigorously lift the mean-variance model to a more general mean-divergence model in which stock returns are allowed to obey any exponential family of distributions. We then devise a general on-line learning algorithm in this setting. We prove for this algorithm the first lower bounds on the most relevant quantity to be optimized in the framework of Markowitz model: the certainty equivalents. Experiments on four real-world stock markets display its ability to track portfolios whose cumulated returns exceed those of the best stock by orders of magnitude.\n" + }, + { + "author": [ + "Babenko, Boris", + "Verma, Nakul", + "Dollar, Piotr", + "Belongie, Serge" + ], + "paper_title": "Multiple Instance Learning with Manifold Bags", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 81, + "page_to": 88, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In many machine learning applications, labeling every instance of data is burdensome. Multiple Instance Learning (MIL), in which training data is provided in the from of labeled bags rather than labeled instances, is one approach for a more relaxed from of supervised learning. Though much progress has been made in analyzing MIL problems, existing work considers bags that have a finite number of instances. In this paper we argue that in many applications of MIL (e.g. image, audio, e.t.c.) the bags are better modeled as low dimensional manifolds in high dimensional feature space. We show that the geometric structure of such manifold bags affects PAC-learnability. We discuss how a learning algorithm that is designed for finite sized bags can be adapted to learn from manifold bags. Furthermore, we propose a simple heuristic that reduces the memory requirements of such algorithms. Our experiments on real-world data validate our analysis and show that our approach works well.\n" + }, + { + "author": [ + "Jiang, Yi", + "Ren, Jiangtao" + ], + "paper_title": "Eigenvalue Sensitive Feature Selection", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 89, + "page_to": 96, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In recent years, some spectral feature selection methods are proposed to choose those features with high power of preserving sample similarity. However, when there exist lots of irrelevant or noisy features in data, the similarity matrix constructed from all the un-weighted features may be not reliable, which then misleads existing spectral feature selection methods to select 'wrong' features. To solve this problem, we propose that feature importance should be evaluated according to their impacts on similarity matrix, which means features with high impacts on similarity matrix should be chosen as important ones. Since graph Laplacian\\cite{luxbury2007} is defined on the similarity matrix, then the impact of each feature on similarity matrix can be reflected on the change of graph Laplacian, especially on its eigen-system. Based on this point of view, we propose an Eigenvalue Sensitive Criteria (EVSC) for feature selection, which aims at seeking those features with high impact on graph Laplacian's eigenvalues. Empirical analysis demonstrates our proposed method outperfroms some traditional spectral feature selection methods.\n" + }, + { + "author": [ + "Su, Jiang", + "Shirab, Jelber Sayyad", + "Matwin, Stan" + ], + "paper_title": "Large Scale Text Classification using Semi-supervised Multinomial Naive Bayes ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 97, + "page_to": 104, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Numerous semi-supervised learning methods have been proposed to augment Multinomial Naive Bayes (MNB) using unlabeled documents, but their use in practice is often limited due to implementation difficulty, inconsistent prediction perfromance, or high computational cost. In this paper, we propose a new, very simple semi-supervised extension of MNB, called Semi-supervised Frequency Estimate (SFE). Our experiments show that it consistently improves MNB with additional data (labeled or unlabeled) in terms of AUC and accuracy, which is not the case when combining MNB with Expectation Maximization (EM). We attribute this to the fact that SFE consistently produces better conditional log likelihood values than both EM+MNB and MNB in labeled training data.\n" + }, + { + "author": [ + "Cho, KyungHyun", + "Raiko, Tapani", + "Ilin, Alexander" + ], + "paper_title": "Enhanced Gradient and Adaptive Learning Rate for Training Restricted Boltzmann Machines", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 105, + "page_to": 112, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Boltzmann machines are often used as building blocks in greedy learning of deep networks. However, training even a simplified model, known as restricted Boltzmann machine (RBM), can be extremely laborious: Traditional learning algorithms often converge only with the right choice of the learning rate scheduling and the scale of the initial weights. They are also sensitive to specific data representation: An equivalent RBM can be obtained by flipping some bits and changing the weights and biases accordingly, but traditional learning rules are not invariant to such transfromations. Without careful tuning of these training settings, traditional algorithms can easily get stuck at plateaus or even diverge. In this work, we present an enhanced gradient which is derived such that it is invariant to bit-flipping transfromations. We also propose a way to automatically adjust the learning rate by maximizing a local likelihood estimate. Our experiments confirm that the proposed improvements yield more stable training of RBMs.\n" + }, + { + "author": [ + "Tarlow, Daniel", + "Batra, Dhruv", + "Kohli, Pushmeet", + "Kolmogorov, Vladimir" + ], + "paper_title": "Dynamic Tree Block Coordinate Ascent", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 113, + "page_to": 120, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:This paper proposes a novel Linear Programming (LP) based algorithm, called Dynamic Tree-Block Coordinate Ascent (DTBCA), for perfroming maximum a posteriori (MAP) inference in probabilistic graphical models. Unlike traditional message passing algorithms, which operate unifromly on the whole factor graph, our method dynamically chooses regions of the factor graph on which to focus message-passing efforts. We propose two criteria for selecting regions, including an efficiently computable upper-bound on the increase in the objective possible by passing messages in any particular region. This bound is derived from the theory of primal-dual methods from combinatorial optimization, and the forest that maximizes the bounds can be chosen efficiently using a maximum-spanning-tree-like algorithm. Experimental results show that our dynamic schedules significantly speed up state-of-the- art LP-based message-passing algorithms on a wide variety of real-world problems.\n" + }, + { + "author": [ + "Mahoney, Michael", + "Orecchia, Lorenzo" + ], + "paper_title": "Implementing regularization implicitly via approximate eigenvector computation", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 121, + "page_to": 128, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Regularization is a powerful technique for extracting useful infromation from noisy data. Typically, it is implemented by adding some sort of norm constraint to an objective function and then exactly optimizing the modified objective function. This procedure often leads to optimization problems that are computationally more expensive than the original problem, a fact that is clearly problematic if one is interested in large-scale applications. On the other hand, a large body of empirical work has demonstrated that heuristics, and in some cases approximation algorithms, developed to speed up computations sometimes have the side-effect of perfroming regularization implicitly. Thus, we consider the question: What is the regularized optimization objective that an approximation algorithm is exactly optimizing? We address this question in the context of computing approximations to the smallest nontrivial eigenvector of a graph Laplacian; and we consider three random-walk-based procedures: one based on the heat kernel of the graph, one based on computing the the PageRank vector associated with the graph, and one based on a truncated lazy random walk. In each case, we provide a precise characterization of the manner in which the approximation method can be viewed as implicitly computing the exact solution to a regularized problem. Interestingly, the regularization is not on the usual vector from of the optimization problem, but instead it is on a related semidefinite program.\n" + }, + { + "author": [ + "Socher, Richard", + "Lin, Cliff Chiung-Yu", + "Ng, Andrew", + "Manning, Chris" + ], + "paper_title": "Parsing Natural Scenes and Natural Language with Recursive Neural Networks", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 129, + "page_to": 136, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Recursive structure is commonly found in the inputs of different modalities such as natural scene images or natural language sentences. Discovering this recursive structure helps us to not only identify the units that an image or sentence contains but also how they interact to from a whole. We introduce a max-margin structure prediction architecture based on recursive neural networks that can successfully recover such structure both in complex scene images as well as sentences. The same algorithm can be used both to provide a competitive syntactic parser for natural language sentences from the Penn Treebank and to outperfrom alternative approaches for semantic scene segmentation, annotation and classification. For segmentation and annotation our algorithm obtains a new level of state-of-the-art perfromance on the Stanford background dataset (78.1%). The features from the image parse tree outperfrom Gist descriptors for scene classification by 4%.\n" + }, + { + "author": [ + "Thomas, Philip", + "Barto, Andrew" + ], + "paper_title": "Conjugate Markov Decision Processes ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 137, + "page_to": 144, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Many open problems involve the search for a mapping that is used by an algorithm solving an MDP. Useful mappings are often from the state set to some other set. Examples include representation discovery (a mapping to a feature space) and skill discovery (a mapping to skill termination probabilities). Different mappings result in algorithms achieving varying expected returns. In this paper we present a novel approach to the search for any mapping used by any algorithm attempting to solve an MDP, for that which results in maximum expected return.\n" + }, + { + "author": [ + "Lu, Tyler", + "Boutilier, Craig" + ], + "paper_title": "Learning Mallows Models with Pairwise Preferences", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 145, + "page_to": 152, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Learning preference distributions is a key problem in many areas (e.g., recommender systems, IR, social choice). However, existing methods require restrictive data models for evidence about user preferences. We relax these restrictions by considering as data arbitrary pairwise comparisons---the fundamental building blocks of ordinal rankings. We develop the first algorithms for learning Mallows models (and mixtures) with pairwise comparisons. At the heart is a new algorithm, the generalized repeated insertion model (GRIM), for sampling from arbitrary ranking distributions. We develop approximate samplers that are exact for many important special cases---and have provable bounds with pairwise evidence---and derive algorithms for evaluating log-likelihood, learning Mallows mixtures, and non-parametric estimation. Experiments on large, real-world datasets show the effectiveness of our approach.\n" + }, + { + "author": [ + "Scott, Clayton" + ], + "paper_title": "Surrogate losses and regret bounds for cost-sensitive classification with example-dependent costs ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 153, + "page_to": 160, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We study surrogate losses in the context of cost-sensitive classification with example-dependent costs, a problem also known as regression level set estimation. We give sufficient conditions on the surrogate loss for the existence of a surrogate regret bound. Such bounds imply that as the surrogate risk tends to its optimal value, so too does the expected misclassification cost. Our sufficient conditions encompass example-dependent versions of the hinge, exponential, and other common losses. These results provide theoretical justification for some previously proposed surrogate-based algorithms, and suggests others that have not yet been developed.\n" + }, + { + "author": [ + "Jawanpuria, Pratik", + "Jagarlapudi, Saketha Nath", + "Ramakrishnan, Ganesh" + ], + "paper_title": "Efficient Rule Ensemble Learning using Hierarchical Kernels ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 161, + "page_to": 168, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:This paper addresses the problem of Rule Ensemble Learning (REL), where the goal is simultaneous discovery of a small set of simple rules and their optimal weights that lead to good generalization. Rules are assumed to be conjunctions of basic propositions concerning the values taken by the input features. From the perspectives of interpretability as well as generalization, it is highly desirable to construct rule ensembles with low training error, having rules that are i) simple, {\\em i.e.}, involve few conjunctions and ii) few in number. We propose to explore the (exponentially) large feature space of all possible conjunctions optimally and efficiently by employing the recently introduced Hierarchical Kernel Learning (HKL) framework. The regularizer employed in the HKL fromulation can be interpreted as a potential for discouraging selection of rules involving large number of conjunctions -- justifying its suitability for constructing rule ensembles. Simulation results show that the proposed approach improves over state-of-the-art REL algorithms in terms of generalization and indeed learns simple rules. Although this is encouraging, it can be shown that HKL selects a conjunction only if all its subsets are selected, and this is highly undesirable. We propose a novel convex fromulation which alleviates this problem and generalizes the HKL framework. The main technical contribution of this paper is an efficient mirror-descent based active set algorithm for solving the new fromulation. Empirical evaluations on REL problems illustrate the utility of generalized HKL.\n" + }, + { + "author": [ + "Martins, Andre", + "Figueiredo, Mario", + "Aguiar, Pedro", + "Smith, Noah", + "Xing, Eric" + ], + "paper_title": "An Augmented Lagrangian Approach to Constrained MAP Inference ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 169, + "page_to": 176, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose a new fast algorithm for approximate MAP inference on factor graphs, which combines augmented Lagrangian optimization with the dual decomposition method. Each slave subproblem is given a quadratic penalty, which pushes toward faster consensus than in previous subgradient approaches. Our algorithm is provably convergent, parallelizable, and suitable for fine decompositions of the graph. We show how it can efficiently handle problems with (possibly global) structural constraints via simple sort operations. Experiments on synthetic and real-world data show that our approach compares favorably with the state-of-the-art.\n" + }, + { + "author": [ + "Mannor, Shie", + "Tsitsiklis, John" + ], + "paper_title": "Mean-Variance Optimization in Markov Decision Processes ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 177, + "page_to": 184, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We consider finite horizon Markov decision processes under perfromance measures that involve both the mean and the variance of the cumulative reward. We show that either randomized or history-based policies can improve perfromance. We prove that the complexity of computing a policy that maximizes the mean reward under a variance constraint is NP-hard for some cases, and strongly NP-hard for others. We finally offer pseudo-polynomial exact and approximation algorithms.\n" + }, + { + "author": [ + "Li, Lei", + "Prakash, B. Aditya" + ], + "paper_title": "Time Series Clustering: Complex is Simpler!", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 185, + "page_to": 192, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Given a motion capture sequence, how to identify the category of the motion? Classifying human motions is a critical task in motion editing and synthesizing, for which manual labeling is clearly inefficient for large databases. Here we study the general problem of time series clustering. We propose a novel method of clustering time series that can (a) learn joint temporal dynamics in the data; (b) handle time lags; and (c) produce interpretable features. We achieve this by developing complex-valued linear dynamical systems (CLDS), which include real-valued Kalman filters as a special case; our advantage is that the transition matrix is simpler (just diagonal), and the transmission one easier to interpret. We then present Complex-Fit, a novel EM algorithm to learn the parameters for the general model and its special case for clustering. Our approach produces significant improvement in clustering quality, 1.5 to 5 times better than well-known competitors on real motion capture sequences.\n" + }, + { + "author": [ + "Gould, Stephen" + ], + "paper_title": "Max-margin Learning for Lower Linear Envelope Potentials in Binary Markov Random Fields", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 193, + "page_to": 200, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The standard approach to max-margin parameter learning for Markov random fields (MRFs) involves incrementally adding the most violated constraints during each iteration of the algorithm. This requires exact MAP inference, which is intractable for many classes of MRF. In this paper, we propose an exact MAP inference algorithm for binary MRFs containing a class of higher-order models, known as lower linear envelope potentials. Our algorithm is polynomial in the number of variables and number of linear envelope functions. With tractable inference in hand, we show how the parameters and corresponding feature vectors can be represented in a max-margin framework for efficiently learning lower linear envelope potentials.\n" + }, + { + "author": [ + "Clark, Alexander" + ], + "paper_title": "Inference of Inversion Transduction Grammars ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 201, + "page_to": 208, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present the first polynomial algorithm for learning a class of inversion transduction grammars (ITGs) that implement context free transducers -- functions from strings to strings. The class of transductions that we can learn properly includes all subsequential transductions. These algorithms are based on a generalisation of distributional learning; we prove correctness of our algorithm under an identification in the limit model.\n" + }, + { + "author": [ + "Hu, En-Liang", + "Wang, Bo", + "Chen, SongCan" + ], + "paper_title": "BCDNPKL: Scalable Non-Parametric Kernel Learning Using Block Coordinate Descent", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 209, + "page_to": 216, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Most existing approaches for non-parametric kernel learning (NPKL) suffer from expensive computation, which would limit their applications to large-scale problems. To address the scalability problem of NPKL, we propose a novel algorithm called BCDNPKL, which is very efficient and scalable. Superior to most existing approaches, BCDNPKL keeps away from semidefinite programming (SDP) and eigen-decomposition, which benefits from two findings: 1) The original SDP framework of NPKL can be reduced into a far smaller-sized counterpart which is corresponding to the sub-kernel (referred to as boundary kernel) learning; 2) The sub-kernel learning can be efficiently solved by using the proposed block coordinate descent (BCD) technique. We provide a fromal proof of global convergence for the proposed BCDNPKL algorithm. The extensive experiments verify the scalability and effectiveness of BCDNPKL, compared with the state-of-the-art algorithms.\n" + }, + { + "author": [ + "der Maaten, Laurens Van" + ], + "paper_title": "Learning Discriminative Fisher Kernels", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 217, + "page_to": 224, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Fisher kernels provide a commonly used vectorial representation of structured objects. The paper presents a technique that exploits label infromation to improve the object representation of Fisher kernels by employing ideas from metric learning. In particular, the new technique trains a generative model in such a way that the distance between the log-likelihood gradients induced by two objects with the same label is as small as possible, and the distance between the gradients induced by two objects with different labels is as large as possible. We illustrate the strong perfromance of classifiers trained on the resulting object representations on problems in handwriting recognition, speech recognition, facial expression analysis, and bio-infromatics.\n" + }, + { + "author": [ + "Kpotufe, Samory", + "von Luxburg, Ulrike" + ], + "paper_title": "Pruning nearest neighbor cluster trees ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 225, + "page_to": 232, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Nearest neighbor ($k$-NN) graphs are widely used in machine learning and data mining applications, and our aim is to better understand what they reveal about the cluster structure of the unknown underlying distribution of points. Moreover, is it possible to identify spurious structures that might arise due to sampling variability? Our first contribution is a statistical analysis that reveals how certain subgraphs of a $k$-NN graph from a consistent estimator of the cluster tree of the underlying distribution of points. Our second and perhaps most important contribution is the following finite sample guarantee. We carefully work out the tradeoff between aggressive and conservative pruning and are able to guarantee the removal of all spurious cluster structures while at the same time guaranteeing the recovery of salient clusters. This is the first such finite sample result in the context of clustering.\n" + }, + { + "author": [ + "Zhao, Peilin", + "Hoi, Steven", + "Jin, Rong", + "Yang, Tianbao" + ], + "paper_title": "Online AUC Maximization ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 233, + "page_to": 240, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Most studies of online learning measure the perfromance of a learner by classification accuracy, which is inappropriate for applications where the data are unevenly distributed among different classes. We address this limitation by developing online learning algorithm for maximizing Area Under the ROC curve (AUC), a metric that is widely used for measuring the classification perfromance for imbalanced data distributions. The key challenge of online AUC maximization is that it needs to optimize the pairwise loss between two instances from different classes. This is in contrast to the classical setup of online learning where the overall loss is a sum of losses over individual training examples. We address this challenge by exploiting the reservoir sampling technique, and present two algorithms for online AUC maximization with theoretic perfromance guarantee. Extensive experimental studies confirm the effectiveness and the efficiency of the proposed algorithms for maximizing AUC.\n" + }, + { + "author": [ + "Yue, Yisong", + "Joachims, Thorsten" + ], + "paper_title": "Beat the Mean Bandit ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 241, + "page_to": 248, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The Dueling Bandits Problem is an online learning framework in which actions are restricted to noisy comparisons between pairs of strategies (also known as bandits). It models settings where absolute rewards are difficult to elicit but pairwise preferences are readily available. In this paper, we extend the Dueling Bandits Problem to a relaxed setting where preference magnitudes can violate transitivity. We present the first algorithm for this more general Dueling Bandits Problem and provide theoretical guarantees in both the online and the PAC settings. Furthermore, we show that the new algorithm has stronger guarantees than existing results even in the original Dueling Bandits Problem, which we validate empirically.\n" + }, + { + "author": [ + "Orabona, Francesco", + "Jie, Luo" + ], + "paper_title": "Ultra-Fast Optimization Algorithm for Sparse Multi Kernel Learning", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 249, + "page_to": 256, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Many state-of-the-art approaches for Multi Kernel Learning (MKL) struggle at finding a compromise between perfromance, sparsity of the solution and speed of the optimization process. In this paper we look at the MKL problem at the same time from a learning and optimization point of view. So, instead of designing a regularizer and then struggling to find an efficient method to minimize it, we design the regularizer while keeping the optimization algorithm in mind. Hence, we introduce a novel MKL fromulation, which mixes elements of p-norm and elastic-net kind of regularization. We also propose a fast stochastic gradient descent method that solves the novel MKL fromulation. We show theoretically and empirically that our method has 1) state-of-the-art perfromance on many classification tasks; 2) exact sparse solutions with a tunable level of sparsity; 3) a convergence rate bound that depends only logarithmically on the number of kernels used, and is independent of the sparsity required; 4) independence on the particular convex loss function used.\n" + }, + { + "author": [ + "Potetz, Brian" + ], + "paper_title": "Estimating the Bayes Point Using Linear Knapsack Problems ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 257, + "page_to": 264, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:A Bayes Point machine is a binary classifier that approximates the Bayes-optimal classifier by estimating the mean of the posterior distribution of classifier parameters. Past Bayes Point machines have overcome the intractability of this goal by using message passing techniques that approximate the posterior of the classifier parameters as a Gaussian distribution. In this paper, we investigate alternative message passing approaches that do not rely on Gaussian approximation. To make this possible, we introduce a new computational shortcut based on linear multiple-choice knapsack problems that reduces the complexity of computing Bayes Point belief propagation messages from exponential to linear in the number of data features. Empirical tests of our approach show significant improvement in linear classification over both soft-margin SVMs and Expectation Propagation Bayes Point machines for several real-world UCI datasets.\n" + }, + { + "author": [ + "Le, Quoc", + "Ngiam, Jiquan", + "Coates, Adam", + "Lahiri, Ahbik", + "Prochnow, Bobby", + "Ng, Andrew" + ], + "paper_title": "On optimization methods for deep learning ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 265, + "page_to": 272, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The predominant methodology in training deep learning advocates the use of stochastic gradient descent methods (SGDs). Despite its ease of implementation, SGDs are difficult to tune and parallelize. These problems make it challenging to develop, debug and scale up deep learning algorithms with SGDs. In this paper, we show that more sophisticated off-the-shelf optimization methods such as Limited memory BFGS (L-BFGS) and Conjugate gradient (CG) with linesearch can significantly simplify and speed up the process of pretraining deep algorithms. In our experiments, the difference between L-BFGS/CG and SGDs are more pronounced if we consider algorithmic extensions (e.g., sparsity regularization) and hardware extensions (e.g., GPUs or computer clusters). Our experiments with distributed optimization support the use of L-BFGS with locally connected networks and convolutional neural networks. Using L-BFGS, our convolutional network model achieves 0.69\\% on the standard MNIST dataset. This is a state-of-the-art result on MNIST among algorithms that do not use distortions or pretraining.\n" + }, + { + "author": [ + "Crammer, Koby", + "Gentile, Claudio" + ], + "paper_title": "Multiclass Classification with Bandit Feedback using Adaptive Regularization ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 273, + "page_to": 280, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present a new multiclass algorithm in the bandit framework, where after making a prediction, the learning algorithm receives only partial feedback, i.e., a single bit of right-or-wrong, rather then the true label. Our algorithm is based on the second-order Perceptron, and uses upper-confidence bounds to trade off exploration and exploitation. We analyze this algorithm in a partial adversarial setting, where instances are chosen adversarially, while the labels are chosen according to a linear probabilistic model, which is also chosen adversarially. From the theoretical viewpoint, we show a regret of O(\\sqrt{T}\\log(T)), which improves over the current best bounds of O(T^{2/3}) in the fully adversarial setting. We evaluate our algorithm on nine real-world text classification problems, obtaining state-of-the-art results, even compared with non-bandit online algorithms, especially when label noise is introduced.\n" + }, + { + "author": [ + "Helmbold, Dave", + "Long, Phil" + ], + "paper_title": "On the Necessity of Irrelevant Variables ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 281, + "page_to": 288, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:This work explores the effects of relevant and irrelevant boolean variables on the accuracy of classifiers. The analysis uses the assumption that the variables are conditionally independent given the class, and focuses on a natural family of learning algorithms for such sources when the relevant variables have a small advantage over random guessing. The main result is that algorithms relying predominately on irrelevant variables have error probabilities that quickly go to 0 in situations where algorithms that limit the use of irrelevant variables have errors bounded below by a positive constant. We also show that accurate learning is possible even when there are so few examples that one cannot determine with high confidence whether or not any individual variable is relevant.\n" + }, + { + "author": [ + "Barthelm\\'{e}, Simon", + "Chopin, Nicolas" + ], + "paper_title": "ABC-EP: Expectation Propagation for Likelihood-free Bayesian Computation ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 289, + "page_to": 296, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Many statistical models of interest to the natural and social sciences have no tractable likelihood function. Until recently, Bayesian inference for such models was thought infeasible. Pritchard et al. (1999) introduced an algorithm known as ABC, for Approximate Bayesian Computation, that enables Bayesian computation in such models. Despite steady progress since this first breakthrough, such as the adaptation of MCMC and Sequential Monte Carlo techniques to likelihood-free inference, state-of-the art methods remain notoriously hard to use and require enormous computation times. Among other issues, one faces the difficult task of finding appropriate summary statistics for the model, and tuning the algorithm can be time-consuming when little prior infromation is available. We show that Expectation Propagation, a widely successful approximate inference technique, can be adapted to the likelihood-free context. The resulting algorithm does not require summary statistics, is an order of magnitude faster than existing techniques, and remains usable when prior infromation is vague.\n" + }, + { + "author": [ + "Germain, Pascal", + "Lacoste, Alexandre", + "Laviolette, Francois", + "Marchand, Mario", + "Shanian, Sara" + ], + "paper_title": "A PAC-Bayes Sample-compression Approach to Kernel Methods ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 297, + "page_to": 304, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose a PAC-Bayes sample compression approach to kernel methods that can accommodate any bounded similarity function and show that the support vector machine (SVM) classifier is a particular case of a more general class of data-dependent classifiers known as majority votes of sample-compressed classifiers. We provide novel risk bounds for these majority votes and learning algorithms that minimize these bounds.\n" + }, + { + "author": [ + "Tamar, Aviv", + "Castro, Dotan Di", + "Meir, Ron" + ], + "paper_title": "Integrating Partial Model Knowledge in Model Free RL Algorithms", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 305, + "page_to": 312, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In reinforcement learning an agent uses online feedback from the environment and prior knowledge in order to adaptively select an effective policy. Model free approaches address this task by directly mapping external and internal states to actions, while model based methods attempt to construct a model of the environment, followed by a selection of optimal actions based on that model. Given the complementary advantages of both approaches, we suggest a novel algorithm which combines them into a single algorithm, which switches between a model based and a model free mode, depending on the current environmental state and on the status of the agent's knowledge. We prove that such an approach leads to improved perfromance whenever environmental knowledge is available, without compromising perfromance when such knowledge is absent. Numerical simulations demonstrate the effectiveness of the approach and suggest its efficacy in boosting policy gradient learning.\n" + }, + { + "author": [ + "\\'{A}lvaro Barbero", + "Sra, Suvrit" + ], + "paper_title": "Fast Newton-type Methods for Total Variation Regularization ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 313, + "page_to": 320, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Numerous applications in statistics, signal processing, and machine learning regularize using Total Variation (TV) penalties. We study anisotropic (l1-based) TV and also a related l2-norm variant. We consider for both variants associated (1D) proximity operators, which lead to challenging optimization problems. We solve these problems by developing Newton-type methods that outperfrom the state-of-the-art algorithms. More importantly, our 1D-TV algorithms serve as building blocks for solving the harder task of computing 2- (and higher)-dimensional TV proximity. We illustrate the computational benefits of our methods by applying them to several applications: (i) image denoising; (ii) image deconvolution (by plugging in our TV solvers into publicly available software); and (iii) four variants of fused-lasso. The results show large speedups--and to support our claims, we provide software accompanying this paper.\n" + }, + { + "author": [ + "Bradley, Joseph", + "Kyrola, Aapo", + "Bickson, Daniel", + "Guestrin, Carlos" + ], + "paper_title": "Parallel Coordinate Descent for L1-Regularized Loss Minimization", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 321, + "page_to": 328, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose Shotgun, a parallel coordinate descent algorithm for minimizing L1-regularized losses. Though coordinate descent seems inherently sequential, we prove convergence bounds for Shotgun which predict linear speedups, up to a problem-dependent limit. We present a comprehensive empirical study of Shotgun for Lasso and sparse logistic regression. Our theoretical predictions on the potential for parallelism closely match behavior on real data. Shotgun outperfroms other published solvers on a range of large problems, proving to be one of the most scalable algorithms for L1.\n" + }, + { + "author": [ + "Shalev-Shwartz, Shai", + "Gonen, Alon", + "Shamir, Ohad" + ], + "paper_title": "Large-Scale Convex Minimization with a Low-Rank Constraint", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 329, + "page_to": 336, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We address the problem of minimizing a convex function over the space of large matrices with low rank. While this optimization problem is hard in general, we propose an efficient greedy algorithm and derive its fromal approximation guarantees. Each iteration of the algorithm involves (approximately) finding the left and right singular vectors corresponding to the largest singular value of a certain matrix, which can be calculated in linear time. This leads to an algorithm which can scale to large matrices arising in several applications such as matrix completion for collaborative filtering and robust low rank matrix approximation.\n" + }, + { + "author": [ + "Hannah, Lauren", + "Dunson, David" + ], + "paper_title": "Approximate Dynamic Programming for Storage Problems ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 337, + "page_to": 344, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Storage problems are an important subclass of stochastic control problems. This paper presents a new method, approximate dynamic programming for storage, to solve storage problems with continuous, convex decision sets. Unlike other solution procedures, ADPS allows math programming to be used to make decisions each time period, even in the presence of large state variables. We test ADPS on the day ahead wind commitment problem with storage.\n" + }, + { + "author": [ + "Jegelka, Stefanie", + "Bilmes, Jeff" + ], + "paper_title": "Online Submodular Minimization for Combinatorial Structures ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 345, + "page_to": 352, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Most results for online decision problems with structured concepts, such as trees or cuts, assume linear costs. In many settings, however, nonlinear costs are more realistic. Owing to their non-separability, these lead to much harder optimization problems. Going beyond linearity, we address online approximation algorithms for structured concepts that allow the cost to be submodular, i.e., nonseparable. In particular, we show regret bounds for three Hannan-consistent strategies that capture different settings. Our results also tighten a regret bound for unconstrained online submodular minimization.\n" + }, + { + "author": [ + "Norouzi, Mohammad", + "Fleet, David" + ], + "paper_title": "Minimal Loss Hashing for Compact Binary Codes", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 353, + "page_to": 360, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose a method for learning similarity-preserving hash functions that map high-dimensional data onto binary codes. The fromulation is based on structured prediction with latent variables and a hinge-like loss function. It is efficient to train for large datasets, scales well to large code lengths, and outperfroms state-of-the-art methods.\n" + }, + { + "author": [ + "Chen, Bo", + "Polatkan, Gungor", + "Sapiro, Guillermo", + "Dunson, David", + "Carin, Lawrence" + ], + "paper_title": "The Hierarchical Beta Process for Convolutional Factor Analysis and Deep Learning", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 361, + "page_to": 368, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:A convolutional factor-analysis model is developed, with the number of filters (factors) inferred via the beta process (BP) and hierarchical BP, for single-task and multi-task learning, respectively. The computation of the model parameters is implemented within a Bayesian setting, employing Gibbs sampling; we explicitly exploit the convolutional nature of the expansion to accelerate computations. The model is used in a multi-level (“deep”) analysis of general data, with specific results presented for image-processing data sets, e.g., classification.\n" + }, + { + "author": [ + "Guillory, Andrew", + "Bilmes, Jeff" + ], + "paper_title": "Simultaneous Learning and Covering with Adversarial Noise ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 369, + "page_to": 376, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We study simultaneous learning and covering problems: submodular set cover problems that depend on the solution to an active (query) learning problem. The goal is to jointly minimize the cost of both learning and covering. We extend recent work in this setting to allow for a limited amount of adversarial noise. Certain noisy query learning problems are a special case of our problem. Crucial to our analysis is a lemma showing the logical OR of two submodular cover constraints can be reduced to a single submodular set cover constraint. Combined with known results, this new lemma allows for arbitrary monotone circuits of submodular cover constraints to be reduced to a single constraint. As an example practical application, we present a movie recommendation website that minimizes the total cost of learning what the user wants to watch and recommending a set of movies.\n" + }, + { + "author": [ + "Chen, Haojun", + "Dunson, David", + "Carin, Lawrence" + ], + "paper_title": "Topic Modeling with Nonparametric Markov Tree", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 377, + "page_to": 384, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:A new hierarchical tree-based topic model is developed, based on nonparametric Bayesian techniques. The model has two unique attributes: (i) a child node in the tree may have more than one parent, with the goal of eliminating redundant sub-topics deep in the tree; and (ii) parsimonious sub-topics are manifested, by removing redundant usage of words at multiple scales. The depth and width of the tree are unbounded within the prior, with a retrospective sampler employed to adaptively infer the appropriate tree size based upon the corpus under study. Excellent quantitative results are manifested on five standard data sets, and the inferred tree structure is also found to be highly interpretable.\n" + }, + { + "author": [ + "Kuwadekar, Ankit", + "Neville, Jennifer" + ], + "paper_title": "Relational Active Learning for Joint Collective Classification Models", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 385, + "page_to": 392, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In many network domains, labeled data may be costly to acquire---indicating a need for {\\em relational active learning} methods. Recent work has demonstrated that relational model perfromance can be improved by taking network structure into account when choosing instances to label. However, in collective inference settings, {\\em both} model estimation {\\em and} prediction can be improved by acquiring a node's label---since relational models estimate a joint distribution over labels in the network and collective classification methods propagate infromation from labeled training data during prediction. This conflates improvement in learning with improvement in inference, since labeling nodes can reduce inference error without improving the overall quality of the learned model. Here, we use {\\em across-network} classification to separate the effects on learning and prediction, and focus on reduction of learning error. When label propagation is used for learning, we find that labeling based on prediction {\\em certainty} is more effective than labeling based on {\\em uncertainty}. As such, we propose a novel active learning method that combines a network-based {\\em certainty} metric with semi-supervised learning and relational resampling. We evaluate our approach on synthetic and real-world networks and show faster learning compared to several baselines, including the network based method of Bilgic et al. 2010.\n" + }, + { + "author": [ + "Kumar, Abhishek", + "III, Hal Daume" + ], + "paper_title": "A Co-training Approach for Multi-view Spectral Clustering ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 393, + "page_to": 400, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose a spectral clustering algorithm for the multi-view setting where we have access to multiple views of the data, each of which can be independently used for clustering. Our spectral clustering algorithm has a flavor of co-training, which is already a widely used idea in semi-supervised learning. We work on the assumption that the true underlying clustering would assign a point to the same cluster irrespective of the view. Hence, we constrain our approach to only search for the clusterings that agree across the views. Our algorithm does not have any hyperparameters to set, which is a major advantage in unsupervised learning. We empirically compare with a number of baseline methods on synthetic and real-world datasets to show the efficacy of the proposed algorithm.\n" + }, + { + "author": [ + "Harel, Maayan", + "Mannor, Shie" + ], + "paper_title": "Learning from Multiple Outlooks", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 401, + "page_to": 408, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose a novel problem fromulation of learning a single task when the data are provided in different feature spaces. Each such space is called an outlook, and is assumed to contain both labeled and unlabeled data. The objective is to take advantage of the data from all the outlooks to better classify each of the outlooks. We devise an algorithm that computes optimal affine mappings from different outlooks to a target outlook by matching moments of the empirical distributions. We further derive a probabilistic interpretation of the resulting algorithm and a sample complexity bound indicating how many samples are needed to adequately find the mapping. We report the results of extensive experiments on activity recognition tasks that show the value of the proposed approach in boosting perfromance.\n" + }, + { + "author": [ + "Cossalter, Michele", + "Yan, Rong", + "Zheng, Lu" + ], + "paper_title": "Adaptive Kernel Approximation for Large-Scale Non-Linear SVM Prediction", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 409, + "page_to": 416, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The applicability of non-linear support vector machines (SVMs) has been limited in large-scale data collections because of their linear prediction complexity to the size of support vectors. We propose an efficient prediction algorithm with perfromance guarantee for non-linear SVMs, termed AdaptSVM. It can selectively collapse the kernel function computation to a reduced set of support vectors, compensated by an additional correction term that can be easily computed on-line. It also allows adaptive fall-back to original kernel computation based on its estimated variance and maximum error tolerance. In addition to theoretical analysis, we empirically evaluate on multiple large-scale datasets to show that the proposed algorithm can speed up the prediction process up to 10000 times with only <0.5 accuracy loss.\n" + }, + { + "author": [ + "Garcia-Garcia, Dario", + "von Luxburg, Ulrike", + "Santos-Rodr\\'{i}iguez, Ra\\'{u}l" + ], + "paper_title": "Risk-Based Generalizations of f-divergences", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 417, + "page_to": 424, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We derive a generalized notion of f-divergences, called (f,l)-divergences. We show that this generalization enjoys many of the nice properties of f-divergences, although it is a richer family. It also provides alternative definitions of standard divergences in terms of surrogate risks. As a first practical application of this theory, we derive a new estimator for the Kulback-Leibler divergence that we use for clustering sets of vectors.\n" + }, + { + "author": [ + "Quadrianto, Novi", + "Lampert, Christoph" + ], + "paper_title": "Learning Multi-View Neighborhood Preserving Projections ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 425, + "page_to": 432, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We address the problem of metric learning for multi-view data, namely the construction of embedding projections from data in different representations into a shared feature space, such that the Euclidean distance in this space provides a meaningful within-view as well as between-view similarity. Our motivation stems from the problem of cross-media retrieval tasks, where the availability of a joint Euclidean distance function is a prerequisite to allow fast, in particular hashing-based, nearest neighbor queries. We fromulate an objective function that expresses the intuitive concept that matching samples are mapped closely together in the output space, whereas non-matching samples are pushed apart, no matter in which view they are available. The resulting optimization problem is not convex, but it can be decomposed explicitly into a convex and a concave part, thereby allowing efficient optimization using the convex-concave procedure. Experiments on an image retrieval task show that nearest-neighbor based cross-view retrieval is indeed possible, and the proposed technique improves the retrieval accuracy over baseline techniques.\n" + }, + { + "author": [ + "Orabona, Francesco", + "Cesa-Bianchi, Nicol\\`{o}" + ], + "paper_title": "Better Algorithms for Selective Sampling ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 433, + "page_to": 440, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We study online algorithms for selective sampling that use regularized least squares (RLS) as base classifier. These algorithms typically perfrom well in practice, and some of them have fromal guarantees on their mistake and query rates. We refine and extend these guarantees in various ways, proposing algorithmic variants that exhibit better empirical behavior while enjoying perfromance guarantees under much more general conditions. We also show a simple way of coupling a generic gradient-based classifier with a specific RLS-based selective sampler, obtaining hybrid algorithms with combined perfromance guarantees.\n" + }, + { + "author": [ + "Robbiano, Sylvain", + "Cl\\'{e}mencon, St\\'{e}phan" + ], + "paper_title": "Minimax Learning Rates for Bipartite Ranking and Plug-in Rules", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 441, + "page_to": 448, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:While it is now well-known in the standard binary classi\fcation setup, that, under suitable margin assumptions and complexity conditions on the regression function, fast or even super-fast rates (i.e. rates faster than n^(-1/2) or even faster than n^-1) can be achieved by plug-in classi\fers, no result of this nature has been proved yet in the context of bipartite ranking, though akin to that of classi\fcation. It is the main purpose of the present paper to investigate this issue. Viewing bipartite ranking as a nested continuous collection of cost-sensitive classi\fcation problems, we exhibit a global low noise condition under which certain plug-in ranking rules can be shown to achieve fast (but not super-fast) rates, establishing thus minimax upper bounds for the excess of ranking risk.\n" + }, + { + "author": [ + "Jetchev, Nikolay", + "Toussaint, Marc" + ], + "paper_title": "Task Space Retrieval Using Inverse Feedback Control ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 449, + "page_to": 456, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Learning complex skills by repeating and generalizing expert behavior is a fundamental problem in robotics. A common approach is learning from demonstration: given examples of correct motions, learn a policy mapping state to action consistent with the training data. However, the usual approaches do not answer the question of what are appropriate representations to generate motions for specific tasks. Inspired by Inverse Optimal Control, we present a novel method to learn latent costs, imitate and generalize demonstrated behavior, and discover a task relevant motion representation: Task Space Retrieval Using Inverse Feedback Control (TRIC). We use the learned latent costs to create motion with a feedback controller. We tested our method on robot grasping of objects, a challenging high-dimensional task. TRIC learns the important control dimensions for the grasping task from a few example movements and is able to robustly approach and grasp objects in new situations.\n" + }, + { + "author": [ + "Virtanen, Seppo", + "Klami, Arto", + "Kaski, Samuel" + ], + "paper_title": "Bayesian CCA via Group Sparsity", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 457, + "page_to": 464, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Bayesian treatments of Canonical Correlation Analysis (CCA) -type latent variable models have been recently proposed for coping with overfitting in small sample sizes, as well as for producing factorizations of the data sources into correlated and non-shared effects. However, all of the current implementations of Bayesian CCA and its extensions are computationally inefficient for high-dimensional data and, as shown in this paper, break down completely for high-dimensional sources with low sample count. Furthermore, they cannot reliably separate the correlated effects from non-shared ones. We propose a new Bayesian CCA variant that is computationally efficient and works for high-dimensional data, while also learning the factorization more accurately. The improvements are gained by introducing a group sparsity assumption and an improved variational approximation. The method is demonstrated to work well on multi-label prediction tasks and in analyzing brain correlates of naturalistic audio stimulation.\n" + }, + { + "author": [ + "Deisenroth, Marc", + "Rasmussen, Carl" + ], + "paper_title": "PILCO: A Model-Based and Data-Efficient Approach to Policy Search ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 465, + "page_to": 472, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper, we introduce PILCO, a practical, data-efficient model-based policy search method. PILCO reduces model bias, one of the key problems of model-based reinforcement learning, in a principled way. By learning a probabilistic dynamics model and explicitly incorporating model uncertainty into long-term planning, PILCO can cope with very little data and facilitates learning from scratch in only a few trials. Policy evaluation is perfromed in closed from using state-of-the-art approximate inference. Furthermore, policy gradients are computed analytically for policy improvement. We report unprecedented learning efficiency on challenging and high-dimensional control tasks.\n" + }, + { + "author": [ + "Karasuyama, Masayuki", + "Takeuchi, Ichiro" + ], + "paper_title": "Suboptimal Solution Path Algorithm for Support Vector Machine ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 473, + "page_to": 480, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We consider a suboptimal solution path algorithm for the Support Vector Machine. The solution path algorithm is known as an effective tool for solving a sequence of a parametrized optimization problems in machine learning. However, the algorithm needs to keep strict optimality conditions satisfied everywhere on the path. This requirement narrows the applicability of the path algorithm and adversely affects its computational efficiency. In our algorithm, user can specify tolerances to the optimality and control the trade-off between accuracy of the solution and the computational cost. We also show that our suboptimal solutions can be interpreted as the solution of a perturbed optimization problem from the original one, provide some theoretical analyses of our algorithm based on a novel interpretation. The experimental results demonstrate the effectiveness of our algorithm in terms of efficiency and accuracy.\n" + }, + { + "author": [ + "Sun, Yi", + "Gomez, Faustino", + "Ring, Mark", + "Schmidhuber, J\\\"{u}rgen" + ], + "paper_title": "Incremental Basis Construction from Temporal Difference Error ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 481, + "page_to": 488, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In many reinforcement-learning (RL) systems, the value function is approximated as a linear combination of a fixed set of basis functions. Perfromance can be improved by adding to this set. Previous approaches construct a series of basis functions that in sufficient number can eventually represent the value function. In contrast, we show that there is a single, ideal basis function, which can directly represent the value function. Its addition to the set immediately reduces the error to zero---without changing existing weights. Moreover, this ideal basis function is simply the value function that results from replacing the MDP's reward function with its Bellman error. This result suggests a novel method for improving value-function estimation: a primary reinforcement learner estimates its value function using its present basis functions; it then sends its TD error to a secondary learner, which interprets that error as a reward function and estimates the corresponding value function; the resulting value function then becomes the primary learner's new basis function. We present both batch and online versions in combination with incremental basis projection, and demonstrate that the perfromance is superior to existing methods, especially in the case of large discount factors.\n" + }, + { + "author": [ + "Gerrish, Sean", + "Blei, David" + ], + "paper_title": "Predicting Legislative Roll Calls from Text ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 489, + "page_to": 496, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We develop several predictive models linking legislative sentiment to legislative text. Our models, which draw on ideas from ideal point estimation and topic models, predict voting patterns based on the contents of bills and infer the political leanings of legislators. With supervised topics, we provide an exploratory window into how the language of the law is correlated with political support. We also derive approximate posterior inference algorithms based on variational methods. Across 12 years of legislative data, we predict specific voting patterns with high accuracy.\n" + }, + { + "author": [ + "Nakajima, Shinichi", + "Sugiyama, Masashi", + "Babacan, Derin" + ], + "paper_title": "On Bayesian PCA: Automatic Dimensionality Selection and Analytic Solution", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 497, + "page_to": 504, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In probabilistic PCA, the fully Bayesian estimation is computationally intractable. To cope with this problem, two types of approximation schemes were introduced: the partially Bayesian PCA (PB-PCA) where only the latent variables are integrated out, and the variational Bayesian PCA (VB-PCA) where the loading vectors are also integrated out. The VB-PCA was proposed as an improved variant of PB-PCA for enabling automatic dimensionality selection (ADS). In this paper, we investigate whether VB-PCA is really the best choice from the viewpoints of computational efficiency and ADS. We first show that ADS is not the unique feature of VB-PCA---PB-PCA is also actually equipped with ADS. We further show that PB-PCA is more advantageous in computational efficiency than VB-PCA because the global solution of PB-PCA can be computed analytically. However, we also show the negative fact that PB-PCA results in a trivial solution in the empirical Bayesian framework. We next consider a simplified variant of VB-PCA, where the latent variables and loading vectors are assumed to be mutually independent (while the ordinary VB-PCA only requires matrix-wise independence). We show that this simplified VB-PCA is the most advantageous in practice because its empirical Bayes solution experimentally works as well as the original VB-PCA, and its global optimal solution can be computed efficiently in a closed from.\n" + }, + { + "author": [ + "Bylander, Tom" + ], + "paper_title": "Learning Linear Functions with Quadratic and Linear Multiplicative Updates", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 505, + "page_to": 512, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We analyze variations of multiplicative updates for learning linear functions online. These can be described as substituting exponentiation in the Exponentiated Gradient (EG) algorithm with quadratic and linear functions. Both kinds of updates substitute exponentiation with simpler operations and reduce dependence on the parameter that specifies the sum of the weights during learning. In particular, the linear multiplicative update places no restrictions on the sum of the weights, and, under a wide range of conditions, achieves worst-case behavior close to the EG algorithm. We perfrom our analysis for square loss and absolute loss, and for regression and classification. We also describe some experiments showing that the perfromance of our algorithms are comparable to EG and the $p$-norm algorithm.\n" + }, + { + "author": [ + "Glorot, Xavier", + "Bordes, Antoine", + "Bengio, Yoshua" + ], + "paper_title": "Domain Adaptation for Large-Scale Sentiment Classification: A Deep Learning Approach", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 513, + "page_to": 520, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The exponential increase in the availability of online reviews and recommendations makes sentiment classification an interesting topic in academic and industrial research. Reviews can span so many different domains that it is difficult to gather annotated training data for all of them. Hence, this paper studies the problem of domain adaptation for sentiment classifiers, hereby a system is trained on labeled reviews from one source domain but is meant to be deployed on another. We propose a deep learning approach which learns to extract a meaningful representation for each review in an unsupervised fashion. Sentiment classifiers trained with this high-level feature representation clearly outperfrom state-of-the-art methods on a benchmark composed of reviews of 4 types of Amazon products. Furthermore, this method scales well and allowed us to successfully perfrom domain adaptation on a larger industrial-strength dataset of 22 domains.\n" + }, + { + "author": [ + "Kang, Zhuoliang", + "Grauman, Kristen", + "Sha, Fei" + ], + "paper_title": "Learning with Whom to Share in Multi-task Feature Learning ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 521, + "page_to": 528, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In multi-task learning (MTL), multiple tasks are learnt jointly. A major assumption for this paradigm is that all those tasks are indeed related so that the joint training is appropriate and beneficial. In this paper, we study the problem of multi-task learning of shared feature representations among tasks, while simultaneously determining ``with whom'' each task should share. We fromulate the problem as a mixed integer programming and provide an alternating minimization technique to solve the optimization problem of jointly identifying grouping structures and parameters. The algorithm monotonically decreases the objective function and converges to a local optimum. Compared to the standard MTL paradigm where all tasks are in a single group, our algorithm improves its perfromance with statistical significance for three out of the four datasets we have studied. We also demonstrate its advantage over other task grouping techniques investigated in literature.\n" + }, + { + "author": [ + "Reyzin, Lev" + ], + "paper_title": "Boosting on a Budget: Sampling for Feature-Efficient Prediction ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 529, + "page_to": 536, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper, we tackle the problem of feature-efficient prediction: classification using a limited number of features per test example. We show that modifying an ensemble classifier such as AdaBoost, by sampling hypotheses from its final weighted predictor, is well-suited for this task. We further consider an extension of this problem, where the costs of examining the various features can differ from one another, and we give an algorithm for this more general setting. We prove the correctness of our algorithms and derive bounds for the number of samples needed for given error rates. We also experimentally verify the effectiveness of our methods.\n" + }, + { + "author": [ + "Ikonomovska, Elena", + "Gama, João", + "Zenko, Bernard", + "Dzeroski, Saso" + ], + "paper_title": "Speeding-Up Hoeffding-Based Regression Trees With Options ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 537, + "page_to": 544, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Data streams are ubiquitous and have in the last two decades become an important research topic. For their predictive non-parametric analysis, Hoeffding-based trees are often a method of choice, offering a possibility of any-time predictions. However, one of their main problems is the delay in learning progress due to the existence of equally discriminative attributes. Options are a natural way to deal with this problem. Option trees build upon regular trees by adding splitting options in the internal nodes. As such they are known to improve accuracy, stability and reduce ambiguity. In this paper, we present on-line option trees for faster learning on numerical data streams. Our results show that options improve the any-time perfromance of ordinary on-line regression trees, while preserving the interpretable structure of trees and without significantly increasing the computational complexity of the algorithm.\n" + }, + { + "author": [ + "Meyer, Gilles", + "Bonnabel, Silvère", + "Sepulchre, Rodolphe" + ], + "paper_title": "Linear Regression under Fixed-Rank Constraints: A Riemannian Approach", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 545, + "page_to": 552, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper, we tackle the problem of learning a linear regression model whose parameter is a fixed-rank matrix. We study the Riemannian manifold geometry of the set of fixed-rank matrices and develop efficient line-search algorithms. The proposed algorithms have many applications, scale to high-dimensional problems, enjoy local convergence properties and confer a geometric basis to recent contributions on learning fixed-rank matrices. Numerical experiments on benchmarks suggest that the proposed algorithms compete with the state-of-the-art, and that manifold optimization offers a versatile framework for the design of rank-constrained machine learning algorithms.\n" + }, + { + "author": [ + "Luo, Dijun", + "Ding, Chris", + "Nie, Feiping", + "Huang, Heng" + ], + "paper_title": "Cauchy Graph Embedding", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 553, + "page_to": 560, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Laplacian embedding provides a low-dimensional representation for the nodes of a graph where the edge weights denote pairwise similarity among the node objects. It is commonly assumed that the Laplacian embedding results preserve the local topology of the original data on the low-dimensional projected subspaces, i.e., for any pair of graph nodes with large similarity, they should be embedded closely in the embedded space. However, in this paper, we will show that the Laplacian embedding often cannot preserve local topology well as we expected. To enhance the local topology preserving property in graph embedding, we propose a novel Cauchy graph embedding which preserves the similarity relationships of the original data in the embedded space via a new objective. Consequentially the machine learning tasks (such as k Nearest Neighbor type classifications) can be easily conducted on the embedded data with better perfromance. The experimental results on both synthetic and real world benchmark data sets demonstrate the usefulness of this new type of embedding.\n" + }, + { + "author": [ + "Rodriguez, Manuel Gomez", + "Balduzzi, David", + "Sch\\\"{o}lkopf, Bernhard" + ], + "paper_title": "Uncovering the Temporal Dynamics of Diffusion Networks ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 561, + "page_to": 568, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Time plays an essential role in the diffusion of infromation, influence and disease over networks. In many cases we only observe when a node copies infromation, makes a decision or becomes infected -- but the connectivity, transmission rates between nodes and transmission sources are unknown. Inferring the underlying dynamics is of outstanding interest since it enables forecasting, influencing and retarding infections, broadly construed. To this end, we model diffusion processes as discrete networks of continuous temporal processes occurring at different rates. Given cascade data -- observed infection times of nodes -- we infer the edges of the global diffusion network and estimate the transmission rates of each edge that best explain the observed data. The optimization problem is convex. The model naturally (without heuristics) imposes sparse solutions and requires no parameter tuning. The problem decouples into a collection of independent smaller problems, thus scaling easily to networks on the order of hundreds of thousands of nodes. Experiments on real and synthetic data show that our algorithm both recovers the edges of diffusion networks and accurately estimates their transmission rates from cascade data.\n" + }, + { + "author": [ + "Gao, Tianshi", + "Koller, Daphne" + ], + "paper_title": "Multiclass Boosting with Hinge Loss based on Output Coding ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 569, + "page_to": 576, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Multiclass classification is an important and fundamental problem in machine learning. A popular family of multiclass classification methods belongs to reducing multiclass to binary based on output coding. Several multiclass boosting algorithms have been proposed to learn the coding matrix and the associated binary classifiers in a problem-dependent way. These algorithms can be unified under a sum-of-exponential loss function defined in the domain of margins. Instead, multiclass SVM uses another type of loss function based on hinge loss. In this paper, we present a new output-coding-based multiclass boosting algorithm using the multiclass hinge loss, which we call HingeBoost.OC. HingeBoost.OC is tested on various real world datasets and shows better perfromance than the existing multiclass boosting algorithm AdaBoost.ERP, one-vs-one, one-vs-all, ECOC and multiclass SVM in a majority of different cases.\n" + }, + { + "author": [ + "Jegelka, Stefanie", + "Bilmes, Jeff" + ], + "paper_title": "Approximation Bounds for Inference using Cooperative Cuts ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 577, + "page_to": 584, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We analyze a family of probability distributions that are characterized by an embedded combinatorial structure. This family includes models having arbitrary treewidth and arbitrary sized factors. Unlike general models with such freedom, where the “most probable explanation” (MPE) problem is inapproximable, the combinatorial structure within our model, in particular the indirect use of submodularity, leads to several MPE algorithms that all have approximation guarantees.\n" + }, + { + "author": [ + "Hernandez-Orallo, Jose", + "Flach, Peter", + "Ferri, Cèsar" + ], + "paper_title": "Brier Curves: a New Cost-Based Visualisation of Classifier Perfromance", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 585, + "page_to": 592, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:It is often necessary to evaluate classifier perfromance over a range of operating conditions, rather than as a point estimate. This is typically assessed through the construction of ‘curves’over a ‘space’, visualising how one or two perfromance metrics vary with the operating condition. For binary classifiers in particular, cost space is a natural way of showing this range of perfromance, visualising loss against operating condition. However, the curves which have been traditionally drawn in cost space, known as cost curves, show the optimal loss, and hence assume knowledge of the optimal decision threshold for a given operating condition. Clearly, this leads to an optimistic assessment of classifier perfromance. In this paper we propose a more natural way of visualising classifier perfromance in cost space, which is to plot probabilistic loss on the y-axis, i.e., the loss arising from the probability estimates. This new curve provides new ways of understanding classifier perfromance and new tools to compare classifiers. In addition, we show that the area under this curve is exactly the Brier score, one of the most popular perfromance metrics for probabilistic classifiers.\n" + }, + { + "author": [ + "Brouard, C\\'{e}line", + "D'Alche-Buc, Florence", + "Szafranski, Marie" + ], + "paper_title": "Semi-supervised Penalized Output Kernel Regression for Link Prediction ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 593, + "page_to": 600, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Link prediction is addressed as an output kernel learning task through semi-supervised Output Kernel Regression. Working in the framework of RKHS theory with vector-valued functions, we establish a new representer theorem devoted to semi-supervised least square regression. We then apply it to get a new model (POKR: Penalized Output Kernel Regression) and show its relevance using numerical experiments on artificial networks and two real applications using a very low percentage of labeled data in a transductive setting.\n" + }, + { + "author": [ + "Nikolenko, Sergey", + "Sirotkin, Alexander" + ], + "paper_title": "A New Bayesian Rating System for Team Competitions ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 601, + "page_to": 608, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present a novel probabilistic rating system for team competitions. Building upon TrueSkill(tm), we change the factor graph structure to cope with the problems of TrueSkill(tm), e.g., multiway ties and variable team size. We give detailed inference algorithms for the new structure. Experimental results show a significant improvement over TrueSkill(tm).\n" + }, + { + "author": [ + "Sujeeth, Arvind", + "Lee, HyoukJoong", + "Brown, Kevin", + "Rompf, Tiark", + "Chafi, Hassan", + "Wu, Michael", + "Atreya, Anand", + "Odersky, Martin", + "Olukotun, Kunle" + ], + "paper_title": "OptiML: An Implicitly Parallel Domain-Specific Language for Machine Learning ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 609, + "page_to": 616, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:As the size of datasets continues to grow, machine learning applications are becoming increasingly limited by the amount of available computational power. Taking advantage of modern hardware requires using multiple parallel programming models targeted at different devices (e.g. CPUs and GPUs). However, programming these devices to run efficiently and correctly is difficult, error-prone, and results in software that is harder to read and maintain. We present OptiML, a domain-specific language (DSL) for machine learning. OptiML is an implicitly parallel, expressive and high perfromance alternative to MATLAB and C++. OptiML perfroms domain-specific analyses and optimizations and automatically generates CUDA code for GPUs. We show that OptiML outperfroms explicitly parallelized MATLAB code in nearly all cases.\n" + }, + { + "author": [ + "Zhu, Jun", + "Chen, Ning", + "Xing, Eric" + ], + "paper_title": "Infinite SVM: a Dirichlet Process Mixture of Large-margin Kernel Machines ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 617, + "page_to": 624, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present Infinite SVM (iSVM), a Dirichlet process mixture of large-margin kernel machines for multi-way classification. An iSVM enjoys the advantages of both Bayesian nonparametrics in handling the unknown number of mixing components, and large-margin kernel machines in robustly capturing local nonlinearity of complex data. We develop an efficient variational learning algorithm for posterior inference of iSVM, and we demonstrate the advantages of iSVM over Dirichlet process mixture of generalized linear models and other benchmarks on both synthetic and real Flickr image classification datasets.\n" + }, + { + "author": [ + "Li, Lingbo", + "Zhou, Mingyuan", + "Sapiro, Guillermo", + "Carin, Lawrence" + ], + "paper_title": "On the Integration of Topic Modeling and Dictionary Learning ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 625, + "page_to": 632, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:A new nonparametric Bayesian model is developed to integrate dictionary learning and topic model into a unified framework. The model is employed to analyze partially annotated images, with the dictionary learning perfromed directly on image patches. Efficient inference is perfromed with a Gibbs-slice sampler, and encouraging results are reported on widely used datasets.\n" + }, + { + "author": [ + "Benjamin Marlin, University of British Columbia", + "Mohammad Khan, University of British Columbia", + "Kevin Murphy, University of British Columbia" + ], + "paper_title": "Piecewise Bounds for Estimating Bernoulli-Logistic Latent Gaussian Models ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 633, + "page_to": 640, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Bernoulli-logistic latent Gaussian models (bLGMs) are a useful model class, but accurate parameter estimation is complicated by the fact that the marginal likelihood contains an intractable logistic-Gaussian integral. In this work, we propose the use of fixed piecewise linear and quadratic upper bounds to the logistic-log-partition (LLP) function as a way of circumventing this intractable integral. We describe a framework for approximately computing minimax optimal piecewise quadratic bounds, as well a generalized expectation maximization algorithm based on using piecewise bounds to estimate bLGMs. We prove a theoretical result relating the maximum error in the LLP bound to the maximum error in the marginal likelihood estimate. Finally, we present empirical results showing that piecewise bounds can be significantly more accurate than previously proposed variational bounds.\n" + }, + { + "author": [ + "Urner, Ruth", + "Shalev-Shwartz, Shai", + "Ben-David, Shai" + ], + "paper_title": "Access to Unlabeled Data can Speed up Prediction Time", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 641, + "page_to": 648, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Semi-supervised learning (SSL) addresses the problem of training a classifier using a small number of labeled examples and many unlabeled examples. Most previous work on SSL focused on how availability of unlabeled data can improve the accuracy of the learned classifiers. In this work we study how unlabeled data can be beneficial for constructing faster classifiers. We propose an SSL algorithmic framework which can utilize unlabeled examples for learning classifiers from a predefined set of fast classifiers. We fromally analyze conditions under which our algorithmic paradigm obtains significant improvements by the use of unlabeled data. As a side benefit of our analysis we propose a novel quantitative measure of the so-called cluster assumption. We demonstrate the potential merits of our approach by conducting experiments on the MNIST data set, showing that, when a sufficiently large unlabeled sample is available, a fast classifier can be learned from much fewer labeled examples than without such a sample.\n" + }, + { + "author": [ + "Roy, Jean-Francis", + "Laviolette, Francois", + "Marchand, Mario" + ], + "paper_title": "From PAC-Bayes Bounds to Quadratic Programs for Majority Votes ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 649, + "page_to": 656, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose to construct a weighted majority vote on a set of basis functions by minimizing a risk bound (called the C-bound) that depends on the first two moments of the margin of the Q-convex combination realized on the training data. This bound minimization algorithm turns out to be a quadratic program that can be efficiently solved. A first version of the algorithm is designed for the supervised inductive setting and turns out to be competitive with AdaBoost, MDBoost and the SVM. The second version of the algorithm, designed for the transductive setting, competes well with TSVM. We also propose a new PAC-Bayes theorem that bounds the difference between the \"true\" value of the C-bound and its empirical estimate and that, unexpectedly, contains no KL-divergence.\n" + }, + { + "author": [ + "Flach, Peter", + "Hernandez-Orallo, Jose", + "Ferri, C\\`{e}sar" + ], + "paper_title": "A Coherent Interpretation of AUC as a Measure of Aggregated Classification Perfromance ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 657, + "page_to": 664, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The area under the ROC curve (AUC), a well-known measure of ranking perfromance, is also often used as a measure of classification perfromance, aggregating over decision thresholds as well as class and cost skews. However, David Hand has recently argued that AUC is fundamentally incoherent as a measure of aggregated classifier perfromance and proposed an alternative measure. Specifically, Hand derives a linear relationship between AUC and expected minimum loss, where the expectation is taken over a distribution of the misclassification cost parameter that depends on the model under consideration. Replacing this distribution with a Beta(2;2) distribution, Hand derives his alternative measure H. In this paper we offer an alternative, coherent interpretation of AUC as linearly related to expected loss. We use a distribution over cost parameter and a distribution over data points, both unifrom and hence model independent. Should one wish to consider only optimal thresholds, we demonstrate that a simple and more intuitive alternative to Hand’s H measure is already available in the from of the area under the cost curve.\n" + }, + { + "author": [ + "Franc, Vojtech", + "Zien, Alexander", + "Sch\\\"{o}lkopf, Bernhard" + ], + "paper_title": "Support Vector Machines as Probabilistic Models ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 665, + "page_to": 672, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We show how the SVM can be viewed as a maximum likelihood estimate of a class of probabilistic models. This model class can be viewed as a reparametrization of the SVM in a similar vein to the $\\nu$-SVM reparametrizing the original SVM. It is not discriminative, but has a non-unifrom marginal. We illustrate the benefits of this new view by re-deriving and re-investigating two established SVM-related algorithms.\n" + }, + { + "author": [ + "Tamuz, Omer", + "Liu, Ce", + "Belongie, Serge", + "Shamir, Ohad", + "Kalai, Adam" + ], + "paper_title": "Adaptively Learning the Crowd Kernel ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 673, + "page_to": 680, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We introduce an algorithm that, given n objects, learns a similarity matrix over all n^2 pairs, from crowdsourced data *alone*. The algorithm samples responses to adaptively chosen triplet-based relative-similarity queries. Each query has the from \"is object a more similar to b or to c?\" and is chosen to be maximally infromative given the preceding responses. The output is an embedding of the objects into Euclidean space (like MDS); we refer to this as the \"crowd kernel.\" SVMs reveal that the crowd kernel captures prominent and subtle features across a number of domains, such as \"is striped\" among neckties and \"vowel vs. consonant\" among letters.\n" + }, + { + "author": [ + "Welling, Max", + "Teh, Yee Whye" + ], + "paper_title": "Bayesian Learning via Stochastic Gradient Langevin Dynamics", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 681, + "page_to": 688, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper we propose a new framework for learning from large scale datasets based on iterative learning from small mini-batches. By adding the right amount of noise to a standard stochastic gradient optimization algorithm we show that the iterates will converge to samples from the true posterior distribution as we anneal the stepsize. This seamless transition between optimization and Bayesian posterior sampling provides an in-built protection against overfitting. We also propose a practical method for Monte Carlo estimates of posterior statistics which monitors a ``sampling threshold'' and collects samples after it has been surpassed. We apply the method to three models: a mixture of Gaussians, logistic regression and ICA with natural gradients.\n" + }, + { + "author": [ + "Ngiam, Jiquan", + "Khosla, Aditya", + "Kim, Mingyu", + "Nam, Juhan", + "Lee, Honglak", + "Ng, Andrew" + ], + "paper_title": "Multimodal Deep Learning", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 689, + "page_to": 696, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Deep networks have been successfully applied to unsupervised feature learning for single modalities (e.g., text, images or audio). In this work, we propose a novel application of deep networks to learn features over multiple modalities. We present a series of tasks for multimodal learning and show how to train deep networks that learn features to address these tasks. In particular, we demonstrate cross modality feature learning, where better features for one modality (e.g., video) can be learned if multiple modalities (e.g., audio and video) are present at feature learning time. Furthermore, we show how to learn a shared representation between modalities and evaluate it on a unique task, where the classifier is trained with audio-only data but tested with video-only data and vice-versa. Our methods are validated on the CUAVE and AVLetters datasets with an audio-visual speech classification task, demonstrating best published visual speech classification on AVLetters and effective shared representation learning.\n" + }, + { + "author": [ + "Kim, JooSeuk", + "Scott, Clayton" + ], + "paper_title": "On the Robustness of Kernel Density M-Estimators ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 697, + "page_to": 704, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We analyze a method for nonparametric density estimation that exhibits robustness to contamination of the training sample. This method achieves robustness by combining a traditional kernel density estimator (KDE) with ideas from classical M-estimation. The KDE based on a Gaussian kernel is interpreted as a sample mean in the associated reproducing kernel Hilbert space (RKHS). This mean is estimated robustly through the use of a robust loss, yielding the so-called robust kernel density estimator (RKDE). This robust sample mean can be found via a kernelized iteratively re-weighted least squares (IRWLS) algorithm. Our contributions are summarized as follows. First, we present a representer theorem for the RKDE, which gives an insight into the robustness of the RKDE. Second, we provide necessary and sufficient conditions for kernel IRWLS to converge to the global minimizer, in the Gaussian RKHS, of the objective function defining the RKDE. Third, characterize and provide a method for computing the influence function associated with the RKDE. Fourth, we illustrate the robustness of the RKDE through experiments on several data sets.\n" + }, + { + "author": [ + "Rai, Piyush", + "III, Hal Daume" + ], + "paper_title": "Beam Search based MAP Estimates for the Indian Buffet Process ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 705, + "page_to": 712, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Nonparametric latent feature models offer a flexible way to discover the latent features underlying the data, without having to a priori specify their number. The Indian Buffet Process (IBP) is a popular example of such a model. Inference in IBP based models, however, remains a challenge. Sampling techniques such as MCMC can be computationally expensive and can take a long time to converge to the stationary distribution. Variational techniques, although faster than sampling, can be difficult to design, and can still remain slow on large data. In many problems, however, we only seek a maximum a posteriori (MAP) estimate of the latent feature assignment matrix. For such cases, we show that techniques such as beam search can give fast, approximate MAP estimates in the IBP based models. If samples from the posterior are desired, these MAP estimates can also serve as sensible initializers for MCMC based algorithms. Experimental results on a variety of datasets suggest that our algorithms can be a computationally viable alternative to Gibbs sampling, the particle filter, and variational inference based approaches for the IBP, and also perfrom better than other heuristics such as greedy search.\n" + }, + { + "author": [ + "Dekel, Ofer", + "Gilad-Bachrach, Ran", + "Shamir, Ohad", + "Xiao, Lin" + ], + "paper_title": "Optimal Distributed Online Prediction ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 713, + "page_to": 720, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Online prediction methods are typically studied as serial algorithms running on a single processor. In this paper, we present the distributed mini-batch (DMB) framework, a method of converting a serial gradient-based online algorithm into a distributed algorithm, and prove an asymptotically optimal regret bound for smooth convex loss functions and stochastic examples. Our analysis explicitly takes into account communication latencies between computing nodes in a network. We also present robust variants, which are resilient to failures and node heterogeneity in an synchronous distributed environment. Our method can also be used for distributed stochastic optimization, attaining an asymptotically linear speedup. Finally, we empirically demonstrate the merits of our approach on large-scale online prediction problems.\n" + }, + { + "author": [ + "Knowles, David", + "Gael, Jurgen Van", + "Ghahramani, Zoubin" + ], + "paper_title": "Message Passing Algorithms for the Dirichlet Diffusion Tree ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 721, + "page_to": 728, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We demonstrate efficient approximate inference for the Dirichlet Diffusion Tree, a Bayesian nonparametric prior over tree structures. Although DDTs provide a powerful and elegant approach for modeling hierarchies they haven't seen much use to date. One problem is the computational cost of MCMC inference. We provide the first deterministic approximate inference methods for DDT models and show excellent perfromance compared to the MCMC alternative. We present message passing algorithms to approximate the Bayesian model evidence for a specific tree. This is used to drive sequential tree building and greedy search to find optimal tree structures, corresponding to hierarchical clusterings of the data. We demonstrate appropriate observation models for continuous and binary data. The empirical perfromance of our method is very close to the computationally expensive MCMC alternative on a density estimation problem, and significantly outperfroms kernel density estimators.\n" + }, + { + "author": [ + "Peng, Jian", + "Hazan, Tamir", + "McAllester, David", + "Urtasun, Raquel" + ], + "paper_title": "Convex Max-Product over Compact Sets for Protein Folding", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 729, + "page_to": 736, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper we present an approach to inference in graphical models with mixture of discrete and bounded continuous variables. In particular, we extend convex max-product to deal with these hybrid models and derive the conditions under which our approach is guaranteed to produce the MAP assignment. When dealing with continuous variables the messages are functions. We investigate a multi-grid approach which can be viewed as a piecewise constant representation of these messages. While this approach provides the theoretical guarantees it is not very practical. Inspired by this, we further propose a particle convex max-product algorithm that significantly outperfroms existing particle methods in the task of protein folding and perfroms comparable to the state-of-the art while using a smaller amount of prior knowledge.\n" + }, + { + "author": [ + "Chakraborty, Doran", + "Stone, Peter" + ], + "paper_title": "Structure Learning in Ergodic Factored MDPs without Knowledge of the Transition Function's In-Degree ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 737, + "page_to": 744, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:This paper introduces Learn Structure and Exploit RMax (LSE-RMax), a novel model based structure learning algorithm for ergodic factored-state MDPs. Given a planning horizon that satisfies a condition, LSE-RMax provably guarantees a return very close to the optimal return, with a high certainty, without requiring any prior knowledge of the in-degree of the transition function as input. LSE-RMax is fully implemented with a thorough analysis of its sample complexity. We also present empirical results demonstrating its effectiveness compared to prior approaches to the problem.\n" + }, + { + "author": [ + "Hocking, Toby", + "Vert, Jean-Philippe", + "Bach, Francis", + "Joulin, Armand" + ], + "paper_title": "Clusterpath: an Algorithm for Clustering using Convex Fusion Penalties ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 745, + "page_to": 752, + "totle_page": 8, + "language": "en", + "abstract": "Abstract: We present a new clustering algorithm by proposing a convex relaxation of hierarchical clustering, which results in a family of objective functions with a natural geometric interpretation. We give efficient algorithms for calculating the continuous regularization path of solutions, and discuss relative advantages of the parameters. Our method experimentally gives state-of-the-art results similar to spectral clustering for non-convex clusters, and has the added benefit of learning a tree structure from the data.\n" + }, + { + "author": [ + "Shieh, Albert", + "Hashimoto, Tatsunori", + "Airoldi, Edo" + ], + "paper_title": "Tree preserving embedding ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 753, + "page_to": 760, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Visualization techniques for complex data are a workhorse of modern scientific pursuits. The goal of visualization is to embed high dimensional data in a low dimensional space, while preserving structure in the data relevant to exploratory data analysis, such as the existence of clusters. However, existing visualization methods often either entirely fail to preserve clusters in embeddings due to the crowding problem or can only preserve clusters at a single resolution. Here, we develop a new approach to visualization, tree preserving embedding (TPE). Our approach takes advantage of the topological notion of connectedness to provably preserve clusters at all resolutions. Our perfromance guarantee holds for finite samples, which makes TPE a robust method for applications. Our approach suggests new strategies for robust data visualization in practice.\n" + }, + { + "author": [ + "Arora, Raman", + "Gupta, Maya", + "Kapila, Amol", + "Maryam Fazel" + ], + "paper_title": "Clustering by Left-Stochastic Matrix Factorization ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 761, + "page_to": 768, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose clustering samples given their pairwise similarities by factorizing the similarity matrix into the product of a cluster probability matrix and its transpose. We propose a rotation-based algorithm to compute this left-stochastic decomposition (LSD). Theoretical results link the LSD clustering method to a soft kernel k-means clustering, give conditions for when the factorization and clustering are unique, and provide error bounds. Experimental results on simulated and real similarity datasets show that the proposed method reliably provides accurate clusterings.\n" + }, + { + "author": [ + "Liu, Miao", + "Liao, Xuejun", + "Carin, Lawrence" + ], + "paper_title": "The Infinite Regionalized Policy Representation ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 769, + "page_to": 776, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We introduce the infinite regionalized policy presentation (iRPR), as a nonparametric policy for reinforcement learning in partially observable Markov decision processes (POMDPs). The iRPR assumes an unbounded set of decision states a priori, and infers the number of states to represent the policy given the experiences. We propose algorithms for learning the number of decision states while maintaining a proper balance between exploration and exploitation. Convergence analysis is provided, along with perfromance evaluations on benchmark problems.\n" + }, + { + "author": [ + "Wick, Michael", + "Rohanimanesh, Khashayar", + "Bellare, Kedar", + "Culotta, Aron", + "McCallum, Andrew" + ], + "paper_title": "SampleRank: Training Factor Graphs with Atomic Gradients", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 777, + "page_to": 784, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present SampleRank, an alternative to contrastive divergence (CD) for estimating parameters in complex graphical models. SampleRank harnesses a user-provided loss function to distribute stochastic gradients across an MCMC chain. As a result, parameter updates can be computed between arbitrary MCMC states. SampleRank is not only faster than CD, but also achieves better accuracy in practice (up to 23\\% error reduction on noun-phrase coreference).\n" + }, + { + "author": [ + "Zhang, XianXing", + "Dunson, David", + "Carin, Lawrence" + ], + "paper_title": "Tree-Structured Infinite Sparse Factor Model", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 785, + "page_to": 792, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:A new tree-structured multiplicative gamma process (TMGP) is developed, for inferring the depth of a tree-based factor-analysis model. This new model is coupled with the nested Chinese restaurant process, to nonparametrically infer the depth and width (structure) of the tree. In addition to developing the model, theoretical properties of the TMGP are addressed, and a novel MCMC sampler is developed. The structure of the inferred tree is used to learn relationships between high-dimensional data, and the model is also applied to compressive sensing and interpolation of incomplete images.\n" + }, + { + "author": [ + "Vattani, Andrea", + "Chakrabarti, Deepayan", + "Gurevich, Maxim" + ], + "paper_title": "Preserving Personalized Pagerank in Subgraphs", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 793, + "page_to": 800, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Choosing a subgraph that can concisely represent a large real-world graph is useful in many scenarios. The usual strategy employed is to sample nodes so that the induced subgraph matches the original graph’s degree distribution, clustering coefficient, etc., but no attempt is made to preserve fine-grained relationships between nodes, which are vital for applications like clustering, classification, and ranking. In this work, we model such relationships via the notion of Personalized PageRank Value (PPV). We show that induced subgraphs output by current sampling methods do not preserve PPVs, and propose algorithms for creating PPV-preserving subgraphs on any given subset of graph nodes. Experiments on three large real-world graphs show that the subgraphs created by our method improve upon induced subgraphs in terms of preserving PPVs, clustering accuracy, and maintaining basic graph properties.\n" + }, + { + "author": [ + "Xiao, Lin", + "Zhou, Dengyong", + "Wu, Mingrui" + ], + "paper_title": "Hierarchical Classification via Orthogonal Transfer ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 801, + "page_to": 808, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We consider multiclass classification problems where the set of labels are organized hierarchically as a category tree. We associate each node in the tree with a classifier and classify the examples recursively from the root to the leaves. We propose a hierarchical Support Vector Machine (SVM) that encourages the classifier at each node to be different from the classifiers at its ancestors. More specifically, we introduce regularizations that force the normal vector of the classifying hyperplane at each node to be orthogonal to those at its ancestors as much as possible. We establish conditions under which training such a hierarchical SVM is a convex optimization problem, and develop an efficient dual-averaging method for solving it. We evaluate the method on a number of real-world text categorization tasks and obtain state-of-the-art perfromance.\n" + }, + { + "author": [ + "Nickel, Maximilian", + "Tresp, Volker", + "Kriegel, Hans-Peter" + ], + "paper_title": "A Three-Way Model for Collective Learning on Multi-Relational Data", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 809, + "page_to": 816, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Relational learning is becoming increasingly important in many areas of application. Here, we present a novel approach to relational learning based on the factorization of a three-way tensor. We show that unlike other tensor approaches, our method is able to perfrom collective learning via the latent components of the model and provide an efficient algorithm to compute the factorization. We substantiate our theoretical considerations regarding the collective learning capabilities of our model by the means of experiments on both a new dataset and a dataset commonly used in entity resolution. Furthermore, we show on common benchmark datasets that our approach achieves better or on-par results, if compared to current state-of-the-art relational learning solutions, while it is significantly faster to compute.\n" + }, + { + "author": [ + "Neumann, Gerhard" + ], + "paper_title": "Variational Inference for Policy Search in changing situations ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 817, + "page_to": 824, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Many policy search algorithms minimize the Kullback-Leibler (KL) divergence to a certain target distribution in order to fit their policy. The commonly used KL-divergence forces the resulting policy to be 'reward-attracted'. The policy tries to reproduce all positively rewarded experience while negative experience is neglected. However, the KL-divergence is not symmetric and we can also minimize the the reversed KL-divergence, which is typically used in variational inference. The policy now becomes 'cost-averse'. It tries to avoid reproducing any negatively-rewarded experience while maximizing exploration. Due to this 'cost-averseness' of the policy, Variational Inference for Policy Search (VIP) has several interesting properties. It requires no kernel-bandwith nor exploration rate, such settings are determined automatically by the inference. The algorithm meets the perfromance of state-of-the-art methods while being applicable to simultaneously learning in multiple situations. We concentrate on using VIP for policy search in robotics. We apply our algorithm to learn dynamic counterbalancing of different kinds of pushes with a human-like 4-link robot.\n" + }, + { + "author": [ + "Buffoni, David", + "Calauzenes, Cl\\'{e}ment", + "Gallinari, Patrick", + "Usunier, Nicolas" + ], + "paper_title": "Learning Scoring Functions with Order-Preserving Losses and Standardized Supervision ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 825, + "page_to": 832, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We address the problem of designing surrogate losses for learning scoring functions in the context of label ranking. We extend to ranking problems a notion of order preserving losses previously introduced for multiclass classi?cation, and show that these losses lead to consistent fromulations with respect to a family of ranking evaluation metrics. An order-preserving loss can be tailored for a given evaluation metric by appropriately setting some weights depending on this metric and the observed supervision. These weights, called the standard from of the supervision, do not always exist, but we show that previous consistency results for ranking were proved in special cases where they do. We then evaluate a new pairwise loss consistent with the (Normalized) Discounted Cumulative Gain on benchmark datasets.\n" + }, + { + "author": [ + "Rifai, Salah", + "Vincent, Pascal", + "Muller, Xavier", + "Glorot, Xavier", + "Bengio, Yoshua" + ], + "paper_title": "Contractive Auto-Encoders: Explicit Invariance During Feature Extraction ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 833, + "page_to": 840, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present in this paper a novel approach for training deterministic auto-encoders. We show that by adding a well chosen penalty term to the classical reconstruction cost function, we can achieve results that equal or surpass those attained by other regularized auto-encoders as well as denoising auto-encoders on a range of datasets. This penalty term corresponds to the Frobenius norm of the Jacobian matrix of the encoder activations with respect to the input. We show that this penalty term results in a localized space contraction which in turn yields robust features on the activation layer. Furthermore, we show how this penalty term is related to both regularized auto-encoders and denoising auto-encoders and how it can be seen as a link between deterministic and non-deterministic auto-encoders. We find empirically that this penalty helps to carve a representation that better captures the local directions of variation dictated by the data, corresponding to a lower-dimensional non-linear manifold, while being more invariant to the vast majority of directions orthogonal to the manifold. Finally, we show that by using the learned features to initialize an MLP, we achieve state of the art classification error on a range of datasets, surpassing other methods of pre-training.\n" + }, + { + "author": [ + "Lazaro-Gredilla, Miguel", + "Titsias, Michalis" + ], + "paper_title": "Variational Heteroscedastic Gaussian Process Regression", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 841, + "page_to": 848, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Standard Gaussian processes (GPs) model observations' noise as constant throughout input space. This is often a too restrictive assumption, but one that is needed for GP inference to be tractable. In this work we present a non-standard variational approximation that allows accurate inference in heteroscedastic GPs (i.e., under input-dependent noise conditions). Computational cost is roughly twice that of the standard GP, and also scales as O(n^3). Accuracy is verified by comparing with the golden standard MCMC and its effectiveness is illustrated on several synthetic and real datasets of diverse characteristics. An application to volatility forecasting is also considered.\n" + }, + { + "author": [ + "Liu, Qiang", + "Ihler, Alexander" + ], + "paper_title": "Bounding the Partition Function using Holder's Inequality ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 849, + "page_to": 856, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We describe an algorithm for approximate inference in graphical models based on Holder's inequality that provides upper and lower bounds on common summation problems such as computing the partition function or probability of evidence in a graphical model. Our algorithm unifies and extends several existing approaches, including variable elimination techniques such as mini-bucket elimination and variational methods such as tree reweighted belief propagation and conditional entropy decomposition. We show that our method inherits benefits from each approach to provide significantly better bounds on sum-product tasks.\n" + }, + { + "author": [ + "Vu, Duy", + "Asuncion, Arthur", + "Hunter, David", + "Smyth, Padhraic" + ], + "paper_title": "Dynamic Egocentric Models for Citation Networks", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 857, + "page_to": 864, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The analysis of the fromation and evolution of networks over time is of fundamental importance to social science, biology, and many other fields. While longitudinal network data sets are increasingly being recorded at the granularity of individual time-stamped events, most studies only focus on collapsed cross-sectional snapshots of the network. In this paper, we introduce a dynamic egocentric framework that models continuous-time network data using multivariate counting processes. For inference, an efficient partial likelihood approach is used, allowing our methods to scale to large networks. We apply our techniques to various citation networks and demonstrate the predictive power and interpretability of the learned statistical models.\n" + }, + { + "author": [ + "Small, Kevin", + "Wallace, Byron", + "Brodley, Carla", + "Trikalinos, Thomas" + ], + "paper_title": "The Constrained Weight Space SVM: Learning with Ranked Features ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 865, + "page_to": 872, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Applying supervised learning methods to new classification tasks requires domain experts to label sufficient training data for the classifier to achieve acceptable perfromance. It is desirable to mitigate this annotation effort. To this end, a pertinent observation is that instance labels are often an indirect from of supervision; it may be more efficient to impart domain knowledge directly to the model in the from of labeled features. We present a novel algorithm for exploiting such domain knowledge which we call the Constrained Weight Space SVM (CW-SVM). In addition to exploiting binary labeled features, our approach allows domain experts to provide ranked features, and, more generally, to express arbitrary expected relationships between sets of features. Our empirical results show that the CW-SVM outperfroms both baseline supervised learning strategies and previously proposed methods for learning with labeled features.\n" + }, + { + "author": [ + "Chen, Yudong", + "Xu, Huan", + "Caramanis, Constantine", + "Sanghavi, Sujay" + ], + "paper_title": "Robust Matrix Completion and Corrupted Columns ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 873, + "page_to": 880, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:This paper considers the problem of matrix completion, when some number of the columns are arbitrarily corrupted. It is well-known that standard algorithms for matrix completion can return arbitrarily poor results, if even a single column is corrupted. What can be done if a large number, or even a constant fraction of columns are corrupted? In this paper, we study this very problem, and develop an robust and efficient algorithm for its solution. One direct application comes from robust collaborative filtering. Here, some number of users are so-called manipulators, and try to skew the predictions of the algorithm. Significantly, our results hold {\\it without any assumptions on the observed entries of the manipulated columns}.\n" + }, + { + "author": [ + "Geramifard, Alborz", + "Doshi, Finale", + "Redding, Joshua", + "Roy, Nicholas", + "How, Jonathan" + ], + "paper_title": "Online Discovery of Feature Dependencies ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 881, + "page_to": 888, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Online representational expansion techniques have improved the learning speed of existing reinforcement learning (RL) algorithms in low dimensional domains, yet existing online expansion methods do not scale well to high dimensional problems. We conjecture that one of the main difficulties limiting this scaling is that features defined over the full-dimensional state space often generalize poorly. Hence, we introduce incremental Feature Dependency Discovery (iFDD) as a computationally-inexpensive method for representational expansion that can be combined with any online, value-based RL method that uses binary features. Unlike other online expansion techniques, iFDD creates new features in low dimensional subspaces of the full state space where feedback errors persist. We provide convergence and computational complexity guarantees for iFDD, as well as showing empirically that iFDD scales well to high dimensional multi-agent planning domains with hundreds of millions of state-action pairs.\n" + }, + { + "author": [ + "Paisley, John", + "Carin, Lawrence", + "Blei, David" + ], + "paper_title": "Variational Inference for Stick-Breaking Beta Process Priors ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 889, + "page_to": 896, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present a variational Bayesian inference algorithm for the stick-breaking construction of the beta process. We derive an alternate representation of the beta process that is amenable to variational inference, and present a bound relating the truncated beta process to its infinite counterpart. We assess perfromance on two matrix factorization problems, using a non-negative factorization model and a linear-Gaussian model.\n" + }, + { + "author": [ + "Babes, Monica", + "Marivate, Vukosi", + "Littman, Michael", + "Subramanian, Kaushik" + ], + "paper_title": "Apprenticeship Learning About Multiple Intentions", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 897, + "page_to": 904, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper, we apply tools from inverse reinforcement learning (IRL) to the problem of learning from (unlabeled) demonstration trajectories of behavior generated by varying ``intentions'' or objectives. We derive an EM approach that clusters observed trajectories by inferring the objectives for each cluster using any of several possible IRL methods, and then uses the constructed clusters to quickly identify the intent of a trajectory. We show that a natural approach to IRL---a gradient ascent method that modifies reward parameters to maximize the likelihood of the observed trajectories---is successful at quickly identifying unknown reward functions. We demonstrate these ideas in the context of apprenticeship learning by acquiring the preferences of a human driver in a simple highway car simulator.\n" + }, + { + "author": [ + "Sohl-Dickstein, Jascha", + "Battaglino, Peter", + "DeWeese, Michael" + ], + "paper_title": "Minimum Probability Flow Learning", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 905, + "page_to": 912, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Fitting probabilistic models to data is often difficult, due to the general intractability of the partition function and its derivatives. Here we propose a new parameter estimation technique that does not require computing an intractable normalization factor or sampling from the equilibrium distribution of the model. This is achieved by establishing dynamics that would transfrom the observed data distribution into the model distribution, and then setting as the objective the minimization of the KL divergence between the data distribution and the distribution produced by running the dynamics for an infinitesimal time. Score matching, minimum velocity learning, and certain froms of contrastive divergence are shown to be special cases of this learning technique. We demonstrate parameter estimation in Ising models, deep belief networks and an independent component analysis model of natural scenes. In the Ising model case, current state of the art techniques are outperfromed by at least an order of magnitude in learning time, with lower error in recovered coupling parameters.\n" + }, + { + "author": [ + "Doshi, Finale", + "Wingate, David", + "Tenenbaum, Josh", + "Roy, Nicholas" + ], + "paper_title": "Infinite Dynamic Bayesian Networks", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 913, + "page_to": 920, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present the infinite dynamic Bayesian network model (iDBN), a nonparametric, factored state-space model that generalizes dynamic Bayesian networks (DBNs). The iDBN can infer every aspect of a DBN: the number of hidden factors, the number of values each factor can take, and (arbitrarily complex) connections and conditionals between factors and observations. In this way, the iDBN generalizes other nonparametric state space models, which until now generally focused on binary hidden nodes and more restricted connection structures. We show how this new prior allows us to find interesting structure in benchmark tests and on two real-world datasets involving weather data and neural infromation flow networks.\n" + }, + { + "author": [ + "Coates, Adam", + "Ng, Andrew" + ], + "paper_title": "The Importance of Encoding Versus Training with Sparse Coding and Vector Quantization ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 921, + "page_to": 928, + "totle_page": 8, + "language": "en", + "abstract": "Abstract: While vector quantization (VQ) has been applied widely to generate features for visual recognition problems, much recent work has focused on more powerful methods. In particular, sparse coding has emerged as a strong alternative to traditional VQ approaches and has been shown to achieve consistently higher perfromance on benchmark datasets. Both approaches can be split into a training phase, where the system learns a dictionary of basis functions from unlabeled data, and an encoding phase, where the dictionary is used to extract features from new inputs. In this work, we investigate the reasons for the success of sparse coding over VQ by decoupling these phases, allowing us to separate out the contributions of the training and encoding in a controlled way. Through extensive experiments on CIFAR, NORB and Caltech 101 datasets, we compare sparse coding and several other training and encoding schemes, including a from of VQ paired with a soft threshold activation function. Our results show not only that we can use fast VQ algorithms for training without penalty, but that we can just as well use randomly chosen exemplars from the training set. Rather than spend resources on training, we find it is more important to choose a good encoder---which can often be as simple as a feed forward non-linearity. Among our results, we demonstrate state-of-the-art perfromance on both CIFAR and NORB.\n" + }, + { + "author": [ + "Cuturi, Marco" + ], + "paper_title": "Fast Global Alignment Kernels", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 929, + "page_to": 936, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose novel approaches to cast the widely-used family of Dynamic Time Warping (DTW) distances and similarities as positive definite kernels for time series. To this effect, we provide new theoretical insights on the family of Global Alignment kernels introduced by Cuturi et al. (2007) and propose alternative kernels which are both positive definite and faster to compute. We provide experimental evidence that these alternatives are both faster and more efficient in classification tasks than other kernels based on the DTW fromalism.\n" + }, + { + "author": [ + "Bazzani, Loris", + "Freitas, Nando", + "Larochelle, Hugo", + "Murino, Vittorio", + "Ting, Jo-Anne" + ], + "paper_title": "Learning attentional policies for tracking and recognition in video with deep networks ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 937, + "page_to": 944, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose a novel attentional model for simultaneous object tracking and recognition that is driven by gaze data. Motivated by theories of the human perceptual system, the model consists of two interacting pathways: ventral and dorsal. The ventral pathway models object appearance and classification using deep (factored)-restricted Boltzmann machines. At each point in time, the observations consist of retinal images, with decaying resolution toward the periphery of the gaze. The dorsal pathway models the location, orientation, scale and speed of the attended object. The posterior distribution of these states is estimated with particle filtering. Deeper in the dorsal pathway, we encounter an attentional mechanism that learns to control gazes so as to minimize tracking uncertainty. The approach is modular (with each module easily replaceable with more sophisticated algorithms), straightforward to implement, practically efficient, and works well in simple video sequences.\n" + }, + { + "author": [ + "Dauphin, Yann", + "Glorot, Xavier", + "Bengio, Yoshua" + ], + "paper_title": "Large-Scale Learning of Embeddings with Reconstruction Sampling", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 945, + "page_to": 952, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper, we present a novel method to speed up the learning of embeddings for large-scale learning tasks involving very sparse data, as is typically the case for Natural Language Processing tasks. Our speed-up method has been developed in the context of Denoising Auto-encoders, which are trained in a purely unsupervised way to capture the input distribution, and learn embeddings for words and text similar to earlier neural language models. The main contribution is a new method to approximate reconstruction error by a sampling procedure. We show how this approximation can be made to obtain an unbiased estimator of the training criterion, and we show how it can be leveraged to make learning much more computationally efficient. We demonstrate the effectiveness of this method on the Amazon and RCV1 NLP datasets. Instead of reducing vocabulary size to make learning practical, our method allows us to train using very large vocabularies. In particular, reconstruction sampling requires 22x less training time on the full Amazon dataset.\n" + }, + { + "author": [ + "Chen, Minmin", + "Weinberger, Kilian", + "Chen, Yixin" + ], + "paper_title": "Automatic Feature Decomposition for Single View Co-training ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 953, + "page_to": 960, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:One of the most successful semi-supervised learning approaches is co-training for multi-view data. In co-training, one trains two classifiers, one for each view, and uses the most confident predictions of the unlabeled data for the two classifiers to ``teach each other''. In this paper, we extend co-training to learning scenarios without an explicit multi-view representation. Inspired by a theoretical analysis of Balcan et. al (2004), we introduce a novel algorithm that splits the feature space during learning, explicitly to encourage co-training to be successful. We demonstrate the efficacy of our proposed method in a weakly-supervised setting on the challenging Caltech-256 object recognition task, where we improve significantly over previous results by (Bergamo & Torresani, 2010) in almost all training-set size settings.\n" + }, + { + "author": [ + "Shin, Kilho", + "Cuturi, Marco", + "Kuboyama, Tetsuji" + ], + "paper_title": "Mapping kernels for trees ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 961, + "page_to": 968, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We propose a comprehensive survey of tree kernels through the lens of the mapping kernels framework. We argue that most existing tree kernels, as well as many more that are presented for the first time in this paper, fall into a typology of kernels whose seemingly intricate computation can be efficiently factorized to yield polynomial time algorithms. Despite this fact, we argue that a naive implementation of such kernels remains prohibitively expensive to compute. We propose an approach whereby some computations for smaller trees are cached, which speeds up considerably the computation of all these tree kernels. We provide experimental evidence of this fact as well as preliminary results on the perfromance of these kernels.\n" + }, + { + "author": [ + "Machart, Pierre", + "Peel, Thomas", + "Anthoine, Sandrine", + "Ralaivola, Liva", + "Glotin, Herv\\'{e}" + ], + "paper_title": "Stochastic Low-Rank Kernel Learning for Regression ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 969, + "page_to": 976, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We present a novel approach to learn a kernel-based regression function. It is based on the use of conical combinations of data-based parameterized kernels and on a new stochastic convex optimization procedure of which we establish convergence guarantees. The overall learning procedure has the nice properties that a) the learned conical combination is automatically designed to perfrom the regression task at hand and b) the updates implicated by the optimization procedure are quite inexpensive. In order to shed light on the appositeness of our learning strategy, we present empirical results from experiments conducted on various benchmark datasets.\n" + }, + { + "author": [ + "Nagano, Kiyohito", + "Kawahara, Yoshinobu", + "Aihara, Kazuyuki" + ], + "paper_title": "Size-constrained Submodular Minimization through Minimum Norm Base", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 977, + "page_to": 984, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:A number of combinatorial optimization problems in machine learning can be described as the problem of minimizing a submodular function. It is known that the unconstrained submodular minimization problem can be solved in strongly polynomial time. However, additional constraints make the problem intractable in many settings. In this paper, we discuss the submodular minimization under a size constraint, which is NP-hard, and generalizes the densest subgraph problem and the unifrom graph partitioning problem. Because of NP-hardness, it is difficult to compute an optimal solution even for a prescribed size constraint. In our approach, we do not give approximation algorithms. Instead, the proposed algorithm computes optimal solutions for some of possible size constraints in polynomial time. Our algorithm utilizes the basic polyhedral theory associated with submodular functions. Additionally, we evaluate the perfromance of the proposed algorithm through computational experiments.\n" + }, + { + "author": [ + "Ladicky, Lubor", + "Torr, Philip" + ], + "paper_title": "Locally Linear Support Vector Machines ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 985, + "page_to": 992, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Linear support vector machines ({\\sc svm}s) have become popular for solving classification tasks due to their fast and simple online application to large scale data sets. However, many problems are not linearly separable. For these problems kernel-based {\\sc svm}s are often used, but unlike their linear variant they suffer from various drawbacks in terms of computational and memory efficiency. Their response can be represented only as a function of the set of support vectors, which has been experimentally shown to grow linearly with the size of the training set. In this paper we propose a novel locally linear {\\sc svm} classifier with smooth decision boundary and bounded curvature. We show how the functions defining the classifier can be approximated using local codings and show how this model can be optimized in an online fashion by perfroming stochastic gradient descent with the same convergence guarantees as standard gradient descent method for linear {\\sc svm}. Our method achieves comparable perfromance to the state-of-the-art whilst being significantly faster than competing kernel {\\sc svm}s. We generalise this model to locally finite dimensional kernel {\\sc svm}.\n" + }, + { + "author": [ + "Kadri, Hachem", + "Rabaoui, Asma", + "Preux, Philippe", + "Duflos, Emmanuel", + "Rakotomamonjy, Alain" + ], + "paper_title": "Functional Regularized Least Squares Classication with Operator-valued Kernels ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 993, + "page_to": 1000, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Although operator-valued kernels have recently received increasing interest in various machine learning and functional data analysis problems such as multi-task learning or functional regression, little attention has been paid to the understanding of their associated feature spaces. In this paper, we explore the potential of adopting an operator-valued kernel feature space perspective for the analysis of functional data. We then extend the Regularized Least Squares Classification (RLSC) algorithm to cover situations where there are multiple functions per observation. Experiments on a sound recognition problem show that the proposed method outperfroms the classical RLSC algorithm.\n" + }, + { + "author": [ + "Jalali, Ali", + "Chen, Yudong", + "Sanghavi, Sujay", + "Xu, Huan" + ], + "paper_title": "Clustering Partially Observed Graphs via Convex Optimization ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1001, + "page_to": 1008, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:This paper considers the problem of clustering a partially observed unweighted graph -- i.e. one where for some node pairs we know there is an edge between them, for some others we know there is no edge, and for the remaining we do not know whether or not there is an edge. We want to organize the nodes into disjoint clusters so that there is relatively dense (observed) connectivity within clusters, and sparse across clusters. We take a novel yet natural approach to this problem, by focusing on finding the clustering that minimizes the number of \"disagreements\" - i.e. the sum of the number of (observed) missing edges within clusters, and (observed) present edges across clusters. Our algorithm uses convex optimization; its basis is a reduction of disagreement minimization to the problem of recovering an (unknown) low-rank matrix and an (unknown) sparse matrix from their partially observed sum. We show that our algorithm succeeds under certain natural assumptions on the optimal clustering and its disagreements. Our results significantly strengthen existing matrix splitting results for the special case of our clustering problem. Our results directly enhance solutions to the problem of Correlation Clustering with partial observations\n" + }, + { + "author": [ + "Yang, Eunho", + "Ravikumar, Pradeep" + ], + "paper_title": "On the Use of Variational Inference for Learning Discrete Graphical Model", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1009, + "page_to": 1016, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We study the general class of estimators for graphical model structure based on optimizing $\\ell_1$-regularized approximate log-likelihood, where the approximate likelihood uses tractable variational approximations of the partition function. We provide a message-passing algorithm that \\emph{directly} computes the $\\ell_1$ regularized approximate MLE. Further, in the case of certain reweighted entropy approximations to the partition function, we show that surprisingly the $\\ell_1$ regularized approximate MLE estimator has a \\emph{closed-from}, so that we would no longer need to run through many iterations of approximate inference and message-passing. Lastly, we analyze this general class of estimators for graph structure recovery, or its \\emph{sparsistency}, and show that it is indeed sparsistent under certain conditions.\n" + }, + { + "author": [ + "Sutskever, Ilya", + "Martens, James", + "Hinton, Geoffrey" + ], + "paper_title": "Generating Text with Recurrent Neural Networks ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1017, + "page_to": 1024, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Recurrent Neural Networks (RNNs) are very powerful sequence models that do not enjoy widespread use because it is extremely difficult to train them properly. Fortunately, recent advances in Hessian-free optimization have been able to overcome the difficulties associated with training RNNs, making it possible to apply them successfully to challenging sequence problems. In this paper we demonstrate the power of RNNs trained with the new Hessian-Free optimizer (HF) by applying them to character-level language modeling tasks. The standard RNN architecture, while effective, is not ideally suited for such tasks, so we introduce a new RNN variant that uses multiplicative (or ``gated'') connections which allow the current input character to determine the transition matrix from one hidden state vector to the next. After training the multiplicative RNN with the HF optimizer for five days on 8 high-end Graphics Processing Units, we were able to surpass the perfromance of the best previous single method for character-level language modeling -- a hierarchical non-parametric sequence model. To our knowledge this represents the largest recurrent neural network application to date.\n" + }, + { + "author": [ + "Agovic, Amrudin", + "Banerjee, Arindam", + "Chatterje, Snigdhansu" + ], + "paper_title": "Probabilistic Matrix Addition", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1025, + "page_to": 1032, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We introduce Probabilistic Matrix Addition (PMA) for modeling real-valued data matrices by simultaneously capturing covariance structure among rows and among columns. PMA additively combines two latent matrices drawn from two Gaussian Processes respectively over rows and columns. The resulting joint distribution over the observed matrix does not factorize over entries, rows, or columns, and can thus capture intricate dependencies in the matrix. Exact inference in PMA is possible, but involves inversion of large matrices, and can be computationally prohibitive. Efficient approximate inference is possible due to the sparse dependency structure among latent variables. We propose two families of approximate inference algorithms for PMA based on Gibbs sampling and MAP inference. We demonstrate the effectiveness of PMA for missing value prediction and multi-label classification problems.\n" + }, + { + "author": [ + "Martens, James", + "Sutskever, Ilya" + ], + "paper_title": "Learning Recurrent Neural Networks with Hessian-Free Optimization ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1033, + "page_to": 1040, + "totle_page": 8, + "language": "en", + "abstract": "Abstract: In this work we resolve the long-outstanding problem of how to effectively train recurrent neural networks (RNNs) on complex and difficult sequence modeling problems which may contain long-term data dependencies. Utilizing recent advances in the Hessian-free optimization approach \\citep{hf}, together with a novel damping scheme, we successfully train RNNs on two sets of challenging problems. First, a collection of pathological synthetic datasets which are known to be impossible for standard optimization approaches (due to their extremely long-term dependencies), and second, on three natural and highly complex real-world sequence datasets where we find that our method significantly outperfroms the previous state-of-the-art method for training neural sequence models: the Long Short-term Memory approach of \\citet{lstm}. Additionally, we offer a new interpretation of the generalized Gauss-Newton matrix of \\citet{schraudolph} which is used within the HF approach of Martens.\n" + }, + { + "author": [ + "Eisenstein, Jacob", + "Ahmed, Amr", + "Xing, Eric" + ], + "paper_title": "Sparse Additive Generative Models of Text ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1041, + "page_to": 1048, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Generative models of text typically associate a multinomial with every class label or topic. Even in simple models this requires the estimation of thousands of parameters; in multifaceted latent variable models, standard approaches require additional latent ``switching'' variables for every token, complicating inference. In this paper, we propose an alternative generative model for text. The central idea is that each class label or latent topic is endowed with a model of the deviation in log-frequency from a constant background distribution. This approach has two key advantages: we can enforce sparsity to prevent overfitting, and we can combine generative facets through simple addition in log space, avoiding the need for latent switching variables. We demonstrate the applicability of this idea to a range of scenarios: classification, topic modeling, and more complex multifaceted generative models.\n" + }, + { + "author": [ + "Gabillon, Victor", + "Lazaric, Alessandro", + "Ghavamzadeh, Mohammad", + "Scherrer, Bruno" + ], + "paper_title": "Classification-based Policy Iteration with a Critic", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1049, + "page_to": 1056, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper, we study the effect of adding a value function approximation component (critic) to rollout classification-based policy iteration (RCPI) algorithms. The idea is to use a critic to approximate the return after we truncate the rollout trajectories. This allows us to control the bias and variance of the rollout estimates of the action-value function. Therefore, the introduction of a critic can improve the accuracy of the rollout estimates, and as a result, enhance the perfromance of the RCPI algorithm. We present a new RCPI algorithm, called direct policy iteration with critic (DPI-Critic), and provide its finite-sample analysis when the critic is based on the LSTD method. We empirically evaluate the perfromance of DPI-Critic and compare it with DPI and LSPI in two benchmark reinforcement learning problems.\n" + }, + { + "author": [ + "Das, Abhimanyu", + "Kempe, David" + ], + "paper_title": "Submodular meets Spectral: Greedy Algorithms for Subset Selection, Sparse Approximation and Dictionary Selection ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1057, + "page_to": 1064, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We study the problem of selecting a subset of $k$ random variables from a large set, in order to obtain the best linear prediction of another variable of interest. This problem can be viewed in the context of both feature selection and sparse approximation. We analyze the perfromance of widely used greedy heuristics, using insights from the maximization of submodular functions and spectral analysis. We introduce the submodularity ratio as a key quantity to help understand why greedy algorithms perfrom well even when the variables are highly correlated. Using our techniques, we obtain the strongest known approximation guarantees for this problem, both in terms of the submodularity ratio and the smallest $k$-sparse eigenvalue of the covariance matrix. We also analyze greedy algorithms for the dictionary selection problem, and significantly improve the previously known guarantees. Our theoretical analysis is complemented by experiments on real-world and synthetic data sets; the experiments show that the submodular ratio is a stronger predictor of the perfromance of greedy algorithms than other spectral parameters.\n" + }, + { + "author": [ + "Parikh, Ankur", + "Song, Le", + "Xing, Eric" + ], + "paper_title": "A Spectral Algorithm for Latent Tree Graphical Models", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1065, + "page_to": 1072, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Latent variable models are powerful tools for probabilistic modeling, and have been successfully applied to various domains, such as speech analysis and bioinfromatics. However, parameter learning algorithms for latent variable models have predominantly relied on local search heuristics such as expectation maximization (EM). We propose a fast, local-minimum-free spectral algorithm for learning latent variable models with arbitrary tree topologies, and show that the joint distribution of the observed variables can be reconstructed from the marginals of triples of observed variables irrespective of the maximum degree of the tree. We demonstrate the perfromance of our spectral algorithm on synthetic and real datasets; for large training sizes, our algorithm perfroms comparable to or better than EM while being orders of magnitude faster.\n" + }, + { + "author": [ + "Guan, Yue", + "Dy, Jennifer", + "Jordan, Michael" + ], + "paper_title": "A Unified Probabilistic Model for Global and Local Unsupervised Feature Selection", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1073, + "page_to": 1080, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Existing algorithms for joint clustering and feature selection can be categorized as either global or local approaches. Global methods select a single cluster-independent subset of features, whereas local methods select cluster-specific subsets of features. In this paper, we present a unified probabilistic model that can perfrom both global and local feature selection for clustering. Our approach is based on a hierarchical beta-Bernoulli prior combined with a Dirichlet process mixture model. We obtain global or local feature selection by adjusting the variance of the beta prior. We provide a variational inference algorithm for our model. In addition to simultaneously learning the clusters and features, this Bayesian fromulation allows us to learn both the number of clusters and the number of features to retain. Experiments on synthetic and real data show that our unified model can find global and local features and cluster data as well as competing methods of each type.\n" + }, + { + "author": [ + "Li, Yu-Feng", + "Zhou, Zhi-Hua" + ], + "paper_title": "Towards Making Unlabeled Data Never Hurt ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1081, + "page_to": 1088, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:It is usually expected that, when labeled data are limited, the learning perfromance can be improved by exploiting unlabeled data. In many cases, however, the perfromances of current semi-supervised learning approaches may be even worse than purely using the limited labeled data.It is desired to have \\textit{safe} semi-supervised learning approaches which never degenerate learning perfromance by using unlabeled data. In this paper, we focus on semi-supervised support vector machines (S3VMs) and propose S4VMs, i.e., safe S3VMs. Unlike S3VMs which typically aim at approaching an optimal low-density separator, S4VMs try to exploit the candidate low-density separators simultaneously to reduce the risk of identifying a poor separator with unlabeled data. We describe two implementations of S4VMs, and our comprehensive experiments show that the overall perfromance of S4VMs are highly competitive to S3VMs, while in contrast to S3VMs which degenerate perfromance in many cases, S4VMs are never significantly inferior to inductive SVMs.\n" + }, + { + "author": [ + "Saxe, Andrew", + "Koh, Pang Wei", + "Chen, Zhenghao", + "Bhand, Maneesh", + "Suresh, Bipin", + "Ng, Andrew" + ], + "paper_title": "On Random Weights and Unsupervised Feature Learning ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1089, + "page_to": 1096, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Recently two anomalous results in the literature have shown that certain feature learning architectures can yield useful features for object recognition tasks even with untrained, random weights. In this paper we pose the question: why do random weights sometimes do so well? Our answer is that certain convolutional pooling architectures can be inherently frequency selective and translation invariant, even with random weights. Based on this we demonstrate the viability of extremely fast architecture search by u\n" + }, + { + "author": [ + "Dudik, Miroslav", + "Langford, John", + "Li, Lihong" + ], + "paper_title": "Doubly Robust Policy Evaluation and Learning ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1097, + "page_to": 1104, + "totle_page": 8, + "language": "en", + "abstract": "Abstract: We study decision making in environments where the reward is only partially observed, but can be modeled as a function of an action and an observed context. This setting, known as contextual bandits, encompasses a wide variety of applications including health-care policy and Internet advertising. A central task is evaluation of a new policy given historic data consisting of contexts, actions and received rewards. The key challenge is that the past data typically does not faithfully represent proportions of actions taken by a new policy. Previous approaches rely either on models of rewards or models of the past policy. The fromer are plagued by a large bias whereas the latter have a large variance. In this work, we leverage the strength and overcome the weaknesses of the two approaches by applying the \\emph{doubly robust} technique to the problems of policy evaluation and optimization. We prove that this approach yields accurate value estimates when we have \\emph{either} a good (but not necessarily consistent) model of rewards \\emph{or} a good (but not necessarily consistent) model of past policy. Extensive empirical comparison demonstrates that the doubly robust approach unifromly improves over existing techniques, achieving both lower variance in value estimation and better policies. As such, we expect the doubly robust approach to become common practice.\n" + }, + { + "author": [ + "Ngiam, Jiquan", + "Chen, Zhenghao", + "Koh, Pang Wei", + "Ng, Andrew" + ], + "paper_title": "Learning Deep Energy Models", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1105, + "page_to": 1112, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Deep generative models with multiple hidden layers have been shown to be able to learn meaningful and compact representations of data. In this work we propose deep energy models, a class of models that use a deep feedforward neural network to model the energy landscape that defines a probabilistic model. We are able to efficiently train all layers of our model at the same time, allowing the lower layers of the model to adapt to the training of the higher layers, producing better generative models. We evaluate the generative perfromance of our models on natural images and demonstrate that joint training of multiple layers yields qualitative and quantitative improvements over greedy layerwise training. We further generalize our models beyond the commonly used sigmoidal neural networks and show how a deep extension of the product of Student-t distributions model achieves good generative perfromance. Finally, we introduce a discriminative extension of our model and demonstrate that it outperfroms other fully-connected models on object recognition on the NORB dataset.\n" + }, + { + "author": [ + "Kotlowski, Wojciech", + "Dembczynski, Krzysztof", + "Huellermeier, Eyke" + ], + "paper_title": "Bipartite Ranking through Minimization of Univariate Loss", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1113, + "page_to": 1120, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Minimization of the rank loss or, equivalently, maximization of the AUC in bipartite ranking calls for minimizing the number of disagreements between pairs of instances. Since the complexity of this problem is inherently quadratic in the number of training examples, it is tempting to ask how much is actually lost by minimizing a simple univariate loss function, as done by standard classification methods, as a surrogate. In this paper, we first note that minimization of 0/1 loss is not an option, as it may yield an arbitrarily high rank loss. We show, however, that better results can be achieved by means of a weighted (cost-sensitive) version of 0/1 loss. Yet, the real gain is obtained through margin-based loss functions, for which we are able to derive proper bounds, not only for rank risk but, more importantly, also for rank regret. The paper is completed with an experimental study in which we address specific questions raised by our theoretical analysis.\n" + }, + { + "author": [ + "Lee, Sangkyun", + "Wright, Stephen" + ], + "paper_title": "Manifold Identification of Dual Averaging Methods for Regularized Stochastic Online Learning", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1121, + "page_to": 1128, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Iterative methods that take steps in approximate subgradient directions have proved to be useful for stochastic learning problems over large or streaming data sets. When the objective consists of a loss function plus a nonsmooth regularization term, whose purpose is to induce structure (for example, sparsity) in the solution, the solution often lies on a low-dimensional manifold along which the regularizer is smooth. This paper shows that a regularized dual averaging algorithm can identify this manifold with high probability. This observation motivates an algorithmic strategy in which, once a near-optimal manifold is identified, we switch to an algorithm that searches only in this manifold, which typically has much lower intrinsic dimension than the full space, thus converging quickly to a near-optimal point with the desired structure. Computational results are presented to illustrate these claims.\n" + }, + { + "author": [ + "Agarwal, Alekh", + "Negahban, Sahand", + "Wainwright, Martin" + ], + "paper_title": "Noisy matrix decomposition via convex relaxation: Optimal rates in high dimensions ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1129, + "page_to": 1136, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We analyze a class of estimators based on a convex relaxation for solving high-dimensional matrix decomposition problems. The observations are the noisy realizations of the sum of an (approximately) low rank matrix $\\Theta^\\star$ with a second matrix $\\Gamma^\\star$ endowed with a complementary from of low-dimensional structure. We derive a general theorem that gives upper bounds on the Frobenius norm error for an estimate of the pair $(\\Theta^\\star, \\Gamma^\\star)$ obtained by solving a convex optimization problem. We then specialize our general result to two cases that have been studied in the context of robust PCA: low rank plus sparse structure, and low rank plus a column sparse structure. Our theory yields Frobenius norm error bounds for both deterministic and stochastic noise matrices, and in the latter case, they are minimax optimal. The sharpness of our theoretical predictions is also confirmed in numerical simulations.\n" + }, + { + "author": [ + "Vainsencher, Daniel", + "Dekel, Ofer", + "Mannor, Shie" + ], + "paper_title": "Bundle Selling by Online Estimation of Valuation Functions", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1137, + "page_to": 1144, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We consider the problem of online selection of a bundle of items when the cost of each item changes arbitrarily from round to round and the valuation function is initially unknown and revealed only through the noisy values of selected bundles (the bandit feedback setting). We are interested in learning schemes that have a small regret compared to an agent who knows the true valuation function. Since there are exponentially many bundles, further assumptions on the valuation functions are needed. We make the assumption that the valuation function is supermodular and has non-linear interactions that are of low degree in a novel sense. We develop efficient learning algorithms that balance exploration and exploitation to achieve low regret in this setting.\n" + }, + { + "author": [ + "Courville, Aarron", + "Bergstra, James", + "Bengio, Yoshua" + ], + "paper_title": "Unsupervised Models of Images by Spike-and-Slab RBMs", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1145, + "page_to": 1152, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The spike and slab Restricted Boltzmann Machine (RBM) is defined by having both a real valued ``slab'' variable and a binary ``spike'' variable associated with each unit in the hidden layer. In this paper we generalize and extend the spike and slab RBM to include non-zero means of the conditional distribution over the observed variables conditional on the binary spike variables. We also introduce a term, quadratic in the observed data that we exploit to guarantee the all conditionals associated with the model are well defined -- a guarantee that was absent in the original spike and slab RBM. The inclusion of these generalizations improves the perfromance of the spike and slab RBM as a feature learner and achieves competitive perfromance on the CIFAR-10 image classification task. The spike and slab model, when trained in a convolutional configuration, can generate sensible samples that demonstrate that the model has capture the broad statistical structure of natural images.\n" + }, + { + "author": [ + "Kamisetty, Hetunandan", + "Xing, Eric", + "Langmead, Christopher" + ], + "paper_title": "Approximating Correlated Equilibria using Relaxations on the Marginal Polytope", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1153, + "page_to": 1160, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In game theory, a Correlated Equilibrium (CE) is an equilibrium concept that generalizes the more well-known Nash Equilibrium. If the game is represented as a graphical game, the computational complexity of computing an optimum CE is exponential in the tree-width of the graph. In settings where this exact computation is not feasible, it is desirable to approximate the properties of the CE, such as its expected social utility and marginal probabilities. We study outer relaxations of this problem that yield approximate marginal strategies for the players under a variety of utility functions. Results on simulated games and in a real problem involving drug design indicate that our approximations can be highly accurate and can be successfully used when exact computation of CE is infeasible.\n" + }, + { + "author": [ + "Yan, Yan", + "Rosales, Romer", + "Fung, Glenn", + "Dy, Jennifer" + ], + "paper_title": "Active Learning from Crowds ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1161, + "page_to": 1168, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Obtaining labels is expensive or time-consuming, but unlabeled data is often abundant and easy to obtain. Many learning task can profit from intelligently choosing unlabeled instances to be labeled by an oracle also known as active learning, instead of simply labeling all the data or randomly selecting data to be labeled. Supervised learning traditionally relies on an oracle playing the role of a teacher. In the multiple annotator paradigm, an oracle, who knows the ground truth, no longer exists; instead, multiple labelers, with varying expertise, are available for querying. This paradigm posits new challenges to the active learning scenario. We can ask which data sample should be labeled next and which annotator should we query to benefit our learning model the most. In this paper, we develop a probabilistic model for learning from multiple annotators that can also learn the annotator expertise even when their expertise may not be consistently accurate (or inaccurate) across the task domain. In addition, we provide an optimization fromulation that allows us to simultaneously learn the most uncertain sample and the annotator/s to query the labels from for active learning. Our active learning approach combines both intelligently selecting samples to label and learning from expertise among multiple labelers to improve learning perfromance.\n" + }, + { + "author": [ + "Waugh, Kevin", + "Ziebart, Brian", + "Bagnell, Drew" + ], + "paper_title": "Computational Rationalization: The Inverse Equilibrium Problem", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1169, + "page_to": 1176, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Modeling the purposeful behavior of imperfect agents from a small number of observations is a challenging task. When restricted to the single-agent decision-theoretic setting, inverse optimal control techniques assume that observed behavior is an approximately optimal solution to an unknown decision problem. These techniques learn a utility function that explains the example behavior and can then be used to accurately predict or imitate future behavior in similar observed or unobserved situations. In this work, we consider similar tasks in competitive and cooperative multi-agent domains. Here, unlike single-agent settings, a player cannot myopically maximize its reward --- it must speculate on how the other agents may act to influence the game's outcome. Employing the game-theoretic notion of regret and the principle of maximum entropy, we introduce a technique for predicting and generalizing behavior, as well as recovering a reward function in these domains.\n" + }, + { + "author": [ + "Ghavamzadeh, Mohammad", + "Lazaric, Alessandro", + "Munos, Remi", + "Hoffman, Matthew" + ], + "paper_title": "Finite-Sample Analysis of Lasso-TD ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1177, + "page_to": 1184, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:In this paper, we analyze the perfromance of Lasso-TD, a modification of LSTD in which the projection operator is defined as a Lasso problem. We first show that Lasso-TD is guaranteed to have a unique fixed point and its algorithmic implementation coincides with the recently presented LARS-TD and LC-TD methods. We then derive two bounds on the prediction error of Lasso-TD in the Markov design setting, i.e., when the perfromance is evaluated on the same states used by the method. The first bound makes no assumption, but has a slow rate w.r.t. the number of samples. The second bound is under an assumption on the empirical Gram matrix, called the compatibility condition, but has an improved rate and directly relates the prediction error to the sparsity of the value function in the feature space at hand.\n" + }, + { + "author": [ + "Pazis, Jason", + "Parr, Ron" + ], + "paper_title": "Generalized Value Functions for Large Action Sets", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1185, + "page_to": 1192, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:The majority of value function approximation based reinforcement learning algorithms available today, focus on approximating the state (V) or state-action (Q) value function and efficient action selection comes as an afterthought. On the other hand, real-world problems tend to have large action spaces, where evaluating every possible action becomes impractical. This mismatch presents a major obstacle in successfully applying reinforcement learning to real-world problems. In this paper we present a unified view of V and Q functions and arrive at a new space-efficient representation, where action selection can be done exponentially faster, without the use of a model. We then describe how to calculate this new value function efficiently via approximate linear programming and provide experimental results that demonstrate the effectiveness of the proposed approach.\n" + }, + { + "author": [ + "Kulesza, Alex", + "Taskar, Ben" + ], + "paper_title": "k-DPPs: Fixed-Size Determinantal Point Processes ", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1193, + "page_to": 1200, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:We consider estimation methods for the class of continuous-data energy based models (EBMs). Our main result shows that estimating the parameters of an EBM using score matching when the conditional distribution over the visible units is Gaussian corresponds to training a particular from of regularized autoencoder. We show how different Gaussian EBMs lead to different autoencoder architectures, providing deep links between these two families of models. We compare the score matching estimator for the mPoT model, a particular Gaussian EBM, to several other training methods on a variety of tasks including image denoising and unsupervised feature extraction. We show that the regularization function induced by score matching leads to superior classification perfromance relative to a standard autoencoder. We also show that score matching yields classification results that are indistinguishable from better-known stochastic approximation maximum likelihood estimators.\n" + }, + { + "author": [ + "Swersky, Kevin", + "Ranzato, Marc'Aurelio", + "Buchman, David", + "Marlin, Benjamin", + "Freitas, Nando" + ], + "paper_title": "On Autoencoders and Score Matching for Energy Based Models", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1201, + "page_to": 1208, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Boosting is a popular way to derive powerful learners from simpler hypothesis classes. Following previous work (Mason et al., 1999; Friedman, 2000) on general boosting frameworks, we analyze gradient-based descent algorithms for boosting with respect to any convex objective and introduce a new measure of weak learner perfromance into this setting which generalizes existing work. We present the first weak to strong learning guarantees for the existing gradient boosting work for smooth convex objectives, and also demonstrate that this work fails for non-smooth objectives. To address this issue, we present new algorithms which extend this boosting approach to arbitrary convex loss functions and give corresponding weak to strong convergence results. In addition, we demonstrate experimental results that support our analysis and demonstrate the need for the new algorithms we present.\n" + }, + { + "author": [ + "Grubb, Alexander", + "Bagnell, Drew" + ], + "paper_title": "Generalized Boosting Algorithms for Convex Optimization", + "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "year": "2011", + "isbn": "978-1-4503-0619-5", + "page_from": 1209, + "page_to": 1216, + "totle_page": 8, + "language": "en", + "abstract": "Abstract:Determinantal point processes (DPPs) have recently been proposed as models for set selection problems where diversity is preferred. For example, they can be used to select diverse sets of sentences to from document summaries, or to find multiple non-overlapping human poses in an image. However, DPPs conflate the modeling of two distinct characteristics: the size of the set, and its content. For many realistic tasks, the size of the desired set is known up front; e.g., in search we may want to show the user exactly ten results. In these situations the effort spent by DPPs modeling set size is not only wasteful, but actually introduces unwanted bias into the modeling of content. Instead, we propose t\n" + } +] diff --git a/db/seeds.rb b/db/seeds.rb index b53fb7b7..cf05956a 100644 --- a/db/seeds.rb +++ b/db/seeds.rb @@ -1,34 +1,62 @@ require 'factory_girl' +require 'json' + +data = File.read("db/data") +data_json = JSON.parse(data) FactoryGirl.define do factory :paper_record, class: "WritingJournal" do |f| - f.sequence(:paper_title_translations) {|n|{zh_tw: "tw_test #{n}", en: "en_test #{n}" }} - f.sequence(:journal_title_translations) {|n| {zh_tw: "tw_test #{n}", en: "en_test #{n}"}} - f.sequence(:keywords) {|n| "keywords #{n}"} - f.sequence(:abstract) {|n| "abstract #{n}"} - f.sequence(:isbn) {|n| "0714312#{n}#{n}#{n}"} - f.sequence(:year) {|n| "201#{n}"} - f.create_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account - f.update_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account - end + f.sequence(:paper_title_translations) do |n| + { zh_tw: "tw_#{data_json[n]["paper_title"]}", + en: "en_#{data_json[n]["paper_title"]}" } + end + + f.sequence(:journal_title_translations) do |n| + {zh_tw: "tw_#{data_json[n]["booktitle"]}", + en: "en_#{data_json[n]["booktitle"]}"} + end + + f.sequence(:abstract) do |n| + "#{data_json[n]["abstract"]}" + end + + f.sequence(:isbn) do |n| + "#{data_json[n]["isbn"]}" + end + + f.sequence(:year) do |n| + "#{data_json[n]["year"]}" + end + + f.sequence(:authors) do |n| + "#{data_json[n]["author"].map{|m| m.split(",").reverse.join(" ")}.join(",")}" + end + + f.sequence(:form_to_start) do |n| + "#{data_json[n]["page_from"]}" + end + + f.sequence(:form_to_end) do |n| + "#{data_json[n]["page_to"]}" + end + + f.sequence(:total_pages) do |n| + "#{data_json[n]["total_page"]}" + end + + f.sequence(:language) do |n| + "#{data_json[n]["language"]}" + end + + f.sequence(:keywords) do |n| + "#{data_json[n]["abstract"].split[-3..-1].join(",")}" + end - factory :custom_record, class: "WritingJournal" do |f| - f.sequence(:paper_title_translations) {|n|{zh_tw: "tw_test #{n}", en: "en_test #{n}" }} - f.sequence(:keywords) {|n| "keywords #{n}"} - f.sequence(:abstract) {|n| "abstract #{n}"} - f.sequence(:isbn) {|n| "0714312#{n}#{n}#{n}"} - f.sequence(:year) {|n| "201#{n}"} f.create_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account f.update_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account - end + end end - -#product 50 records -# 50.times.each do FactoryGirl.create(:paper_record) end - -FactoryGirl.create(:custom_record, journal_title_translations: {zh_tw: "tw_test A", en: "en_test A"}) -FactoryGirl.create(:custom_record, journal_title_translations: {zh_tw: "tw_test B", en: "en_test B"}) From 065d38c797503793b601159c00edb7439e84a41e Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Mon, 26 Nov 2012 14:44:35 +0800 Subject: [PATCH 02/83] add autocomplete js --- app/controllers/desktop/journal_pages_controller.rb | 7 ++++++- app/views/desktop/journal_pages/_form.html.erb | 3 +++ 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/app/controllers/desktop/journal_pages_controller.rb b/app/controllers/desktop/journal_pages_controller.rb index 6288cdbf..66c613ef 100644 --- a/app/controllers/desktop/journal_pages_controller.rb +++ b/app/controllers/desktop/journal_pages_controller.rb @@ -18,6 +18,11 @@ class Desktop::JournalPagesController < ApplicationController @level_types = JournalLevelType.all @author_types = JournalAuthorType.all @paper_types= JournalPaperType.all + @co_author_candidate = + CoAuthor.where(name_id: current_user.id).map{|c|c.co_author} + @journal_candidate = + WritingJournal.where(create_user_id: current_user.id).map{|j|j.journal_title}.uniq + respond_to do |format| format.html { render :layout => false} @@ -26,7 +31,7 @@ class Desktop::JournalPagesController < ApplicationController end def edit - @writing_journal= WritingJournal.find(params[:id]) + @writing_journal = WritingJournal.find(params[:id]) @level_types = JournalLevelType.all @author_types = JournalAuthorType.all @paper_types= JournalPaperType.all diff --git a/app/views/desktop/journal_pages/_form.html.erb b/app/views/desktop/journal_pages/_form.html.erb index 20fad7a1..b2328dc4 100644 --- a/app/views/desktop/journal_pages/_form.html.erb +++ b/app/views/desktop/journal_pages/_form.html.erb @@ -207,3 +207,6 @@ + From 0125010a0592a27b9c221750e7cdd9a7b5d00609 Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Mon, 26 Nov 2012 15:45:51 +0800 Subject: [PATCH 03/83] modified error msgs and cancle button, but those two have bug... --- .../desktop/co_authors_controller.rb | 6 ++++-- .../desktop/journal_pages_controller.rb | 20 ++++--------------- .../desktop/journal_pages/_form.html.erb | 4 ++-- 3 files changed, 10 insertions(+), 20 deletions(-) diff --git a/app/controllers/desktop/co_authors_controller.rb b/app/controllers/desktop/co_authors_controller.rb index 2ea928bc..d41e2348 100644 --- a/app/controllers/desktop/co_authors_controller.rb +++ b/app/controllers/desktop/co_authors_controller.rb @@ -39,7 +39,8 @@ class Desktop::CoAuthorsController < ApplicationController if @co_author.save render json: {success:true, msg: "Co-author successfully saved!"}.to_json else - render json: {success: false, msg: @co_author.errors.full_messages}.to_json + error_msg = @co_author.errors.full_messages.join("
") + render json: {success: false, msg: error_msg}.to_json end end @@ -49,7 +50,8 @@ class Desktop::CoAuthorsController < ApplicationController if @co_author.update_attributes(params[:co_author]) render json: {success:true, msg: "Co-author successfully update!"}.to_json else - render json: {success: false, msg: @co_author.errors.full_messages}.to_json + error_msg = @co_author.errors.full_messages.join("
") + render json: {success: false, msg: error_msg}.to_json end end diff --git a/app/controllers/desktop/journal_pages_controller.rb b/app/controllers/desktop/journal_pages_controller.rb index 66c613ef..69334a5b 100644 --- a/app/controllers/desktop/journal_pages_controller.rb +++ b/app/controllers/desktop/journal_pages_controller.rb @@ -23,7 +23,6 @@ class Desktop::JournalPagesController < ApplicationController @journal_candidate = WritingJournal.where(create_user_id: current_user.id).map{|j|j.journal_title}.uniq - respond_to do |format| format.html { render :layout => false} end @@ -47,7 +46,8 @@ class Desktop::JournalPagesController < ApplicationController if @writing_journal.save render json: {success: true, msg: "Paper successfully saved!"}.to_json else - render json: {success: false, msg: @writing_journal.errors.full_messages}.to_json + error_msg = @writing_journal.errors.full_messages.join("
") + render json: {success: false, msg: error_msg}.to_json end end @@ -58,7 +58,8 @@ class Desktop::JournalPagesController < ApplicationController if @writing_journal.update_attributes(params[:writing_journal]) render json: {success: true, msg: "Paper successfully saved!"}.to_json else - render json: {success: false, msg: @writing_journal.errors.full_messages}.to_json + error_msg = @writing_journal.errors.full_messages.join("
") + render json: {success: false, msg: error_msg}.to_json end end end @@ -66,19 +67,6 @@ class Desktop::JournalPagesController < ApplicationController def check_file_type file if not file.nil? file_type = MIME::Types.type_for(file).first.to_s.split("/")[1] - - # case file_type - # when "jpg", "jpeg" - # type = "jpg" - # when "text", "txt" - # type = "txt" - # when "pdf" - # type = "pdf" - # when "png" - # type = "png" - # else "readme" - # end - file_type = "/assets/ft-icons/#{file_type}/#{file_type}-48_32.png" else file_type = "" diff --git a/app/views/desktop/journal_pages/_form.html.erb b/app/views/desktop/journal_pages/_form.html.erb index b2328dc4..5f88322e 100644 --- a/app/views/desktop/journal_pages/_form.html.erb +++ b/app/views/desktop/journal_pages/_form.html.erb @@ -1,8 +1,8 @@
<%= f.submit "Save", name: "commit", value: "Save", class: "fn_btn hh2 thmc2 thmtxt" %> - + + <%= submit_tag "Cancel", :type => "button", class: "bt-cancel fn_btn hh2 thmc2 thmtxt" %>
Entry Year
From c9e3d5271a07a33d80ba1987c96bfecf07b8c9fb Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Mon, 26 Nov 2012 16:06:22 +0800 Subject: [PATCH 04/83] remove '"' candidate list --- app/views/desktop/journal_pages/_form.html.erb | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/app/views/desktop/journal_pages/_form.html.erb b/app/views/desktop/journal_pages/_form.html.erb index 5f88322e..334a6eee 100644 --- a/app/views/desktop/journal_pages/_form.html.erb +++ b/app/views/desktop/journal_pages/_form.html.erb @@ -208,5 +208,5 @@
From 6de62318c54669714f0ebcae2d9dfb1f98107ffb Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Tue, 27 Nov 2012 00:05:04 +0800 Subject: [PATCH 05/83] fake data for journal and coauthor --- db/data | 918 ++++++++++++++++++++++++++-------------------------- db/seeds.rb | 123 ++++--- 2 files changed, 533 insertions(+), 508 deletions(-) diff --git a/db/data b/db/data index 98407fac..4d7f6e52 100644 --- a/db/data +++ b/db/data @@ -7,14 +7,14 @@ "Chang, Shih-Fu" ], "paper_title": "Hashing with Graphs ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1, + "page_form": 1, "page_to": 8, "totle_page": 8, "language": "en", - "abstract": "Abstract:Hashing is becoming increasingly popular for efficient nearest neighbor search in massive databases. However, learning short codes that yield good search perfromance is still a challenge. Moreover, in many cases real-world data lives on a low-dimensional manifold, which should be taken into account to capture meaningful nearest neighbors. In this paper, we propose a novel graph-based hashing method which automatically discovers the neighborhood structure inherent in the data to learn appropriate compact codes. To make such an approach computationally feasible, we utilize Anchor Graphs to obtain tractable low-rank adjacency matrices. Our fromulation allows constant time hashing of a new data point by extrapolating graph Laplacian eigenvectors to eigenfunctions. Finally, we describe a hierarchical threshold learning procedure in which each eigenfunction yields multiple bits, leading to higher search accuracy. Experimental comparison with the other state-of-the-art methods on two large datasets demonstrates the efficacy of the proposed method.\n" + "abstract": "Abstract:Hashing is becoming increasingly popular for efficient nearest neighbor search in massive databases. However, learning short codes that yield good search performance is still a challenge. Moreover, in many cases real-world data lives on a low-dimensional manifold, which should be taken into account to capture meaningful nearest neighbors. In this paper, we propose a novel graph-based hashing method which automatically discovers the neighborhood structure inherent in the data to learn appropriate compact codes. To make such an approach computationally feasible, we utilize Anchor Graphs to obtain tractable low-rank adjacency matrices. Our formulation allows constant time hashing of a new data point by extrapolating graph Laplacian eigenvectors to eigenfunctions. Finally, we describe a hierarchical threshold learning procedure in which each eigenfunction yields multiple bits, leading to higher search accuracy. Experimental comparison with the other state-of-the-art methods on two large datasets demonstrates the efficacy of the proposed method." }, { "author": [ @@ -22,14 +22,14 @@ "Kwok, James" ], "paper_title": "Efficient Sparse Modeling with Automatic Feature Grouping", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Neural Computation", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 9, + "page_form": 9, "page_to": 16, "totle_page": 8, "language": "en", - "abstract": "Abstract:The grouping of features is highly beneficial in learning with high-dimensional data. It reduces the variance in the estimation and improves the stability of feature selection, leading to improved generalization. Moreover, it can also help in data understanding and interpretation. OSCAR is a recent sparse modeling tool that achieves this by using a $\\ell_1$-regularizer and a pairwise $\\ell_\\infty$-regularizer. However, its optimization is computationally expensive. In this paper, we propose an efficient solver based on the accelerated gradient methods. We show that its key projection step can be solved by a simple iterative group merging algorithm. It is highly efficient and reduces the empirical time complexity from $O(d^3 \\sim d^5)$ for the existing solvers to just $O(d)$, where $d$ is the number of features. Experimental results on toy and real-world data sets demonstrate that OSCAR is a competitive sparse modeling approach with the added ability of automatic feature grouping.\n" + "abstract": "Abstract:The grouping of features is highly beneficial in learning with high-dimensional data. It reduces the variance in the estimation and improves the stability of feature selection, leading to improved generalization. Moreover, it can also help in data understanding and interpretation. OSCAR is a recent sparse modeling tool that achieves this by using a $\\ell_1$-regularizer and a pairwise $\\ell_\\infty$-regularizer. However, its optimization is computationally expensive. In this paper, we propose an efficient solver based on the accelerated gradient methods. We show that its key projection step can be solved by a simple iterative group merging algorithm. It is highly efficient and reduces the empirical time complexity from $O(d^3 \\sim d^5)$ for the existing solvers to just $O(d)$, where $d$ is the number of features. Experimental results on toy and real-world data sets demonstrate that OSCAR is a competitive sparse modeling approach with the added ability of automatic feature grouping." }, { "author": [ @@ -37,14 +37,14 @@ "Kwok, James" ], "paper_title": "Multi-Label Classification on Tree- and DAG-Structured Hierarchies ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IEEE Transactions on Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 17, + "page_form": 17, "page_to": 24, "totle_page": 8, "language": "en", - "abstract": "Abstract:Many real-world applications involve multi-label classification, in which the labels are organized in the from of a tree or directed acyclic graph (DAG). However, current research efforts typically ignore the label dependencies or can only exploit the dependencies in tree-structured hierarchies. In this paper, we present a novel hierarchical multi-label classification algorithm which can be used on both tree- and DAG-structured hierarchies. The key idea is to fromulate the search for the optimal consistent multi-label as the finding of the best subgraph in a tree/DAG. Using a simple greedy strategy, the proposed algorithm is computationally efficient, easy to implement, does not suffer from the problem of insufficient/skewed training data in classifier training, and can be readily used on large hierarchies. Theoretical results guarantee the optimality of the obtained solution. Experiments are perfromed on a large number of functional genomics data sets. The proposed method consistently outperfroms the state-of-the-art method on both tree- and DAG-structured hierarchies.\n" + "abstract": "Abstract:Many real-world applications involve multi-label classification, in which the labels are organized in the form of a tree or directed acyclic graph (DAG). However, current research efforts typically ignore the label dependencies or can only exploit the dependencies in tree-structured hierarchies. In this paper, we present a novel hierarchical multi-label classification algorithm which can be used on both tree- and DAG-structured hierarchies. The key idea is to formulate the search for the optimal consistent multi-label as the finding of the best subgraph in a tree/DAG. Using a simple greedy strategy, the proposed algorithm is computationally efficient, easy to implement, does not suffer from the problem of insufficient/skewed training data in classifier training, and can be readily used on large hierarchies. Theoretical results guarantee the optimality of the obtained solution. Experiments are performed on a large number of functional genomics data sets. The proposed method consistently outperforms the state-of-the-art method on both tree- and DAG-structured hierarchies." }, { "author": [ @@ -52,14 +52,14 @@ "Lawrence, Rick" ], "paper_title": "A Graph-based Framework for Multi-Task Multi-View Learning ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Machine Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 25, + "page_form": 25, "page_to": 32, "totle_page": 8, "language": "en", - "abstract": "Abstract:Many real-world problems exhibit dual-heterogeneity. A single learning task might have features in multiple views (i.e., feature heterogeneity); multiple learning tasks might be related with each other through one or more shared views (i.e., task heterogeneity). Existing multi-task learning or multi-view learning algorithms only capture one type of heterogeneity. In this paper, we introduce Multi-Task Multi-View (M^2TV) learning for such complicated learning problems with both feature heterogeneity and task heterogeneity. We propose a graph-based framework (GraM^2) to take full advantage of the dual-heterogeneous nature. Our framework has a natural connection to Reproducing Kernel Hilbert Space (RKHS). Furthermore, we propose an iterative algorithm (IteM^2) for GraM^2 framework, and analyze its optimality, convergence and time complexity. Experimental results on various real data sets demonstrate its effectiveness.\n" + "abstract": "Abstract:Many real-world problems exhibit dual-heterogeneity. A single learning task might have features in multiple views (i.e., feature heterogeneity); multiple learning tasks might be related with each other through one or more shared views (i.e., task heterogeneity). Existing multi-task learning or multi-view learning algorithms only capture one type of heterogeneity. In this paper, we introduce Multi-Task Multi-View (M^2TV) learning for such complicated learning problems with both feature heterogeneity and task heterogeneity. We propose a graph-based framework (GraM^2) to take full advantage of the dual-heterogeneous nature. Our framework has a natural connection to Reproducing Kernel Hilbert Space (RKHS). Furthermore, we propose an iterative algorithm (IteM^2) for GraM^2 framework, and analyze its optimality, convergence and time complexity. Experimental results on various real data sets demonstrate its effectiveness." }, { "author": [ @@ -67,14 +67,14 @@ "Tao, Dacheng" ], "paper_title": "GoDec: Randomized Low-rank & Sparse Matrix Decomposition in Noisy Case ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 33, + "page_form": 33, "page_to": 40, "totle_page": 8, "language": "en", - "abstract": "Abstract:Low-rank and sparse structures have been profoundly studied in matrix completion and compressed sensing. In this paper, we develop ``Go Decomposition'' (GoDec) to efficiently and robustly estimate the low-rank part $L$ and the sparse part $S$ of a matrix $X=L+S+G$ with noise $G$. GoDec alternatively assigns the low-rank approximation of $X-S$ to $L$ and the sparse approximation of $X-L$ to $S$. The algorithm can be significantly accelerated by bilateral random projections (BRP). We also propose GoDec for matrix completion as an important variant. We prove that the objective value $\\|X-L-S\\|_F^2$ converges to a local minimum, while $L$ and $S$ linearly converge to local optimums. Theoretically, we analyze the influence of $L$, $S$ and $G$ to the asymptotic/convergence speeds in order to discover the robustness of GoDec. Empirical studies suggest the efficiency, robustness and effectiveness of GoDec comparing with representative matrix decomposition and completion tools, e.g., Robust PCA and OptSpace.\n" + "abstract": "Abstract:Low-rank and sparse structures have been profoundly studied in matrix completion and compressed sensing. In this paper, we develop ``Go Decomposition'' (GoDec) to efficiently and robustly estimate the low-rank part $L$ and the sparse part $S$ of a matrix $X=L+S+G$ with noise $G$. GoDec alternatively assigns the low-rank approximation of $X-S$ to $L$ and the sparse approximation of $X-L$ to $S$. The algorithm can be significantly accelerated by bilateral random projections (BRP). We also propose GoDec for matrix completion as an important variant. We prove that the objective value $\\|X-L-S\\|_F^2$ converges to a local minimum, while $L$ and $S$ linearly converge to local optimums. Theoretically, we analyze the influence of $L$, $S$ and $G$ to the asymptotic/convergence speeds in order to discover the robustness of GoDec. Empirical studies suggest the efficiency, robustness and effectiveness of GoDec comparing with representative matrix decomposition and completion tools, e.g., Robust PCA and OptSpace." }, { "author": [ @@ -82,14 +82,14 @@ "Mannor, Shie" ], "paper_title": "Unimodal Bandits", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Journal of Machine Learning Research", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 41, + "page_form": 41, "page_to": 48, "totle_page": 8, "language": "en", - "abstract": "Abstract:We consider multiarmed bandit problems where the expected reward is unimodal over partially ordered arms. In particular, the arms may belong to a continuous interval or correspond to vertices in a graph, where the graph structure represents similarity in rewards. The unimodality assumption has an important advantage: we can determine if a given arm is optimal by sampling the possible directions around it. This property allows us to quickly and efficiently find the optimal arm and detect abrupt changes in the reward distributions. For the case of bandits on graphs, we incur a regret proportional to the maximal degree and the diameter of the graph, instead of the total number of vertices.\n" + "abstract": "Abstract:We consider multiarmed bandit problems where the expected reward is unimodal over partially ordered arms. In particular, the arms may belong to a continuous interval or correspond to vertices in a graph, where the graph structure represents similarity in rewards. The unimodality assumption has an important advantage: we can determine if a given arm is optimal by sampling the possible directions around it. This property allows us to quickly and efficiently find the optimal arm and detect abrupt changes in the reward distributions. For the case of bandits on graphs, we incur a regret proportional to the maximal degree and the diameter of the graph, instead of the total number of vertices." }, { "author": [ @@ -99,14 +99,14 @@ "Pillonetto, Gianluigi" ], "paper_title": "Learning Output Kernels with Block Coordinate Descent ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Pattern Recognition Letters", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 49, + "page_form": 49, "page_to": 56, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose a method to learn simultaneously a vector-valued function and a kernel between its components. The obtained kernel can be used both to improve learning perfromances and to reveal structures in the output space which may be important in their own right. Our method is based on the solution of a suitable regularization problem over a reproducing kernel Hilbert space (RKHS) of vector-valued functions. Although the regularized risk functional is non-convex, we show that it is invex, implying that all local minimizers are global minimizers. We derive a block-wise coordinate descent method that efficiently exploits the structure of the objective functional. Then, we empirically demonstrate that the proposed method can improve classification accuracy. Finally, we provide a visual interpretation of the learned kernel matrix for some well known datasets.\n" + "abstract": "Abstract:We propose a method to learn simultaneously a vector-valued function and a kernel between its components. The obtained kernel can be used both to improve learning performances and to reveal structures in the output space which may be important in their own right. Our method is based on the solution of a suitable regularization problem over a reproducing kernel Hilbert space (RKHS) of vector-valued functions. Although the regularized risk functional is non-convex, we show that it is invex, implying that all local minimizers are global minimizers. We derive a block-wise coordinate descent method that efficiently exploits the structure of the objective functional. Then, we empirically demonstrate that the proposed method can improve classification accuracy. Finally, we provide a visual interpretation of the learned kernel matrix for some well known datasets." }, { "author": [ @@ -114,14 +114,14 @@ "Sindhwani, Vikas" ], "paper_title": "Vector-valued Manifold Regularization ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "The Journal of Artificial Societies and Social Simulation", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 57, + "page_form": 57, "page_to": 64, "totle_page": 8, "language": "en", - "abstract": "Abstract:We consider the general problem of learning an unknown functional dependency, f : X->Y, between a structured input space X and a structured output space Y, from labeled and unlabeled examples. We fromulate this problem in terms of data-dependent regularization in Vector-valued Reproducing Kernel Hilbert Spaces (Micchelli & Pontil, 2005) which elegantly extend familiar scalar-valued kernel methods to the general setting where Y has a Hilbert space structure. Our methods provide a natural extension of Manifold Regularization (Belkin et al., 2006) algorithms to also exploit output inter-dependencies while enforcing smoothness with respect to input data geometry. We propose a class of matrix-valued kernels which allow efficient implementations of our algorithms via the use of numerical solvers for Sylvester matrix equations. On multilabel image annotation and text classification problems, we find favorable empirical comparisons against several competing alternatives.\n" + "abstract": "Abstract:We consider the general problem of learning an unknown functional dependency, f : X->Y, between a structured input space X and a structured output space Y, from labeled and unlabeled examples. We formulate this problem in terms of data-dependent regularization in Vector-valued Reproducing Kernel Hilbert Spaces (Micchelli & Pontil, 2005) which elegantly extend familiar scalar-valued kernel methods to the general setting where Y has a Hilbert space structure. Our methods provide a natural extension of Manifold Regularization (Belkin et al., 2006) algorithms to also exploit output inter-dependencies while enforcing smoothness with respect to input data geometry. We propose a class of matrix-valued kernels which allow efficient implementations of our algorithms via the use of numerical solvers for Sylvester matrix equations. On multilabel image annotation and text classification problems, we find favorable empirical comparisons against several competing alternatives." }, { "author": [ @@ -130,15 +130,15 @@ "Kimura, Manabu", "Hachiya, Hirotaka" ], - "paper_title": "On Infromation-Maximization Clustering: Tuning Parameter Selection and Analytic Solution ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "paper_title": "On Information-Maximization Clustering: Tuning Parameter Selection and Analytic Solution ", + "booktitle": "International Journal of Pattern Recognition and Artificial Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 65, + "page_form": 65, "page_to": 72, "totle_page": 8, "language": "en", - "abstract": "Abstract:Infromation-maximization clustering learns a probabilistic classifier in an unsupervised manner so that mutual infromation between feature vectors and cluster assignments is maximized. A notable advantage of this approach is that it only involves continuous optimization of model parameters, which is substantially easier to solve than discrete optimization of cluster assignments. However, existing methods still involve non-convex optimization problems, and therefore finding a good local optimal solution is not straightforward in practice. In this paper, we propose an alternative infromation-maximization clustering method based on a squared-loss variant of mutual infromation. This novel approach gives a clustering solution analytically in a computationally efficient way via kernel eigenvalue decomposition. Furthermore, we provide a practical model selection procedure that allows us to objectively optimize tuning parameters included in the kernel function. Through experiments, we demonstrate the usefulness of the proposed approach.\n" + "abstract": "Abstract:Information-maximization clustering learns a probabilistic classifier in an unsupervised manner so that mutual information between feature vectors and cluster assignments is maximized. A notable advantage of this approach is that it only involves continuous optimization of model parameters, which is substantially easier to solve than discrete optimization of cluster assignments. However, existing methods still involve non-convex optimization problems, and therefore finding a good local optimal solution is not straightforward in practice. In this paper, we propose an alternative information-maximization clustering method based on a squared-loss variant of mutual information. This novel approach gives a clustering solution analytically in a computationally efficient way via kernel eigenvalue decomposition. Furthermore, we provide a practical model selection procedure that allows us to objectively optimize tuning parameters included in the kernel function. Through experiments, we demonstrate the usefulness of the proposed approach." }, { "author": [ @@ -148,14 +148,14 @@ "Nielsen, Frank" ], "paper_title": "On tracking portfolios with certainty equivalents on a generalization of Markowitz model: the Fool, the Wise and the Adaptive", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Information Research", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 73, + "page_form": 73, "page_to": 80, "totle_page": 8, "language": "en", - "abstract": "Abstract:Portfolio allocation theory has been heavily influenced by a major contribution of Harry Markowitz in the early fifties: the mean-variance approach. While there has been a continuous line of works in on-line learning portfolios over the past decades, very few works have really tried to cope with Markowitz model. A major drawback of the mean-variance approach is that it is approximation-free only when stock returns obey a Gaussian distribution, an assumption known not to hold in real data. In this paper, we first alleviate this assumption, and rigorously lift the mean-variance model to a more general mean-divergence model in which stock returns are allowed to obey any exponential family of distributions. We then devise a general on-line learning algorithm in this setting. We prove for this algorithm the first lower bounds on the most relevant quantity to be optimized in the framework of Markowitz model: the certainty equivalents. Experiments on four real-world stock markets display its ability to track portfolios whose cumulated returns exceed those of the best stock by orders of magnitude.\n" + "abstract": "Abstract:Portfolio allocation theory has been heavily influenced by a major contribution of Harry Markowitz in the early fifties: the mean-variance approach. While there has been a continuous line of works in on-line learning portfolios over the past decades, very few works have really tried to cope with Markowitz model. A major drawback of the mean-variance approach is that it is approximation-free only when stock returns obey a Gaussian distribution, an assumption known not to hold in real data. In this paper, we first alleviate this assumption, and rigorously lift the mean-variance model to a more general mean-divergence model in which stock returns are allowed to obey any exponential family of distributions. We then devise a general on-line learning algorithm in this setting. We prove for this algorithm the first lower bounds on the most relevant quantity to be optimized in the framework of Markowitz model: the certainty equivalents. Experiments on four real-world stock markets display its ability to track portfolios whose cumulated returns exceed those of the best stock by orders of magnitude." }, { "author": [ @@ -165,14 +165,14 @@ "Belongie, Serge" ], "paper_title": "Multiple Instance Learning with Manifold Bags", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Advanced Robotics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 81, + "page_form": 81, "page_to": 88, "totle_page": 8, "language": "en", - "abstract": "Abstract:In many machine learning applications, labeling every instance of data is burdensome. Multiple Instance Learning (MIL), in which training data is provided in the from of labeled bags rather than labeled instances, is one approach for a more relaxed from of supervised learning. Though much progress has been made in analyzing MIL problems, existing work considers bags that have a finite number of instances. In this paper we argue that in many applications of MIL (e.g. image, audio, e.t.c.) the bags are better modeled as low dimensional manifolds in high dimensional feature space. We show that the geometric structure of such manifold bags affects PAC-learnability. We discuss how a learning algorithm that is designed for finite sized bags can be adapted to learn from manifold bags. Furthermore, we propose a simple heuristic that reduces the memory requirements of such algorithms. Our experiments on real-world data validate our analysis and show that our approach works well.\n" + "abstract": "Abstract:In many machine learning applications, labeling every instance of data is burdensome. Multiple Instance Learning (MIL), in which training data is provided in the form of labeled bags rather than labeled instances, is one approach for a more relaxed form of supervised learning. Though much progress has been made in analyzing MIL problems, existing work considers bags that have a finite number of instances. In this paper we argue that in many applications of MIL (e.g. image, audio, e.t.c.) the bags are better modeled as low dimensional manifolds in high dimensional feature space. We show that the geometric structure of such manifold bags affects PAC-learnability. We discuss how a learning algorithm that is designed for finite sized bags can be adapted to learn from manifold bags. Furthermore, we propose a simple heuristic that reduces the memory requirements of such algorithms. Our experiments on real-world data validate our analysis and show that our approach works well." }, { "author": [ @@ -180,14 +180,14 @@ "Ren, Jiangtao" ], "paper_title": "Eigenvalue Sensitive Feature Selection", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Information Systems Frontiers", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 89, + "page_form": 89, "page_to": 96, "totle_page": 8, "language": "en", - "abstract": "Abstract:In recent years, some spectral feature selection methods are proposed to choose those features with high power of preserving sample similarity. However, when there exist lots of irrelevant or noisy features in data, the similarity matrix constructed from all the un-weighted features may be not reliable, which then misleads existing spectral feature selection methods to select 'wrong' features. To solve this problem, we propose that feature importance should be evaluated according to their impacts on similarity matrix, which means features with high impacts on similarity matrix should be chosen as important ones. Since graph Laplacian\\cite{luxbury2007} is defined on the similarity matrix, then the impact of each feature on similarity matrix can be reflected on the change of graph Laplacian, especially on its eigen-system. Based on this point of view, we propose an Eigenvalue Sensitive Criteria (EVSC) for feature selection, which aims at seeking those features with high impact on graph Laplacian's eigenvalues. Empirical analysis demonstrates our proposed method outperfroms some traditional spectral feature selection methods.\n" + "abstract": "Abstract:In recent years, some spectral feature selection methods are proposed to choose those features with high power of preserving sample similarity. However, when there exist lots of irrelevant or noisy features in data, the similarity matrix constructed from all the un-weighted features may be not reliable, which then misleads existing spectral feature selection methods to select 'wrong' features. To solve this problem, we propose that feature importance should be evaluated according to their impacts on similarity matrix, which means features with high impacts on similarity matrix should be chosen as important ones. Since graph Laplacian\\cite{luxbury2007} is defined on the similarity matrix, then the impact of each feature on similarity matrix can be reflected on the change of graph Laplacian, especially on its eigen-system. Based on this point of view, we propose an Eigenvalue Sensitive Criteria (EVSC) for feature selection, which aims at seeking those features with high impact on graph Laplacian's eigenvalues. Empirical analysis demonstrates our proposed method outperforms some traditional spectral feature selection methods." }, { "author": [ @@ -196,14 +196,14 @@ "Matwin, Stan" ], "paper_title": "Large Scale Text Classification using Semi-supervised Multinomial Naive Bayes ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Pattern Analysis and Applications", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 97, + "page_form": 97, "page_to": 104, "totle_page": 8, "language": "en", - "abstract": "Abstract:Numerous semi-supervised learning methods have been proposed to augment Multinomial Naive Bayes (MNB) using unlabeled documents, but their use in practice is often limited due to implementation difficulty, inconsistent prediction perfromance, or high computational cost. In this paper, we propose a new, very simple semi-supervised extension of MNB, called Semi-supervised Frequency Estimate (SFE). Our experiments show that it consistently improves MNB with additional data (labeled or unlabeled) in terms of AUC and accuracy, which is not the case when combining MNB with Expectation Maximization (EM). We attribute this to the fact that SFE consistently produces better conditional log likelihood values than both EM+MNB and MNB in labeled training data.\n" + "abstract": "Abstract:Numerous semi-supervised learning methods have been proposed to augment Multinomial Naive Bayes (MNB) using unlabeled documents, but their use in practice is often limited due to implementation difficulty, inconsistent prediction performance, or high computational cost. In this paper, we propose a new, very simple semi-supervised extension of MNB, called Semi-supervised Frequency Estimate (SFE). Our experiments show that it consistently improves MNB with additional data (labeled or unlabeled) in terms of AUC and accuracy, which is not the case when combining MNB with Expectation Maximization (EM). We attribute this to the fact that SFE consistently produces better conditional log likelihood values than both EM+MNB and MNB in labeled training data." }, { "author": [ @@ -212,14 +212,14 @@ "Ilin, Alexander" ], "paper_title": "Enhanced Gradient and Adaptive Learning Rate for Training Restricted Boltzmann Machines", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal on Document Analysis and Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 105, + "page_form": 105, "page_to": 112, "totle_page": 8, "language": "en", - "abstract": "Abstract:Boltzmann machines are often used as building blocks in greedy learning of deep networks. However, training even a simplified model, known as restricted Boltzmann machine (RBM), can be extremely laborious: Traditional learning algorithms often converge only with the right choice of the learning rate scheduling and the scale of the initial weights. They are also sensitive to specific data representation: An equivalent RBM can be obtained by flipping some bits and changing the weights and biases accordingly, but traditional learning rules are not invariant to such transfromations. Without careful tuning of these training settings, traditional algorithms can easily get stuck at plateaus or even diverge. In this work, we present an enhanced gradient which is derived such that it is invariant to bit-flipping transfromations. We also propose a way to automatically adjust the learning rate by maximizing a local likelihood estimate. Our experiments confirm that the proposed improvements yield more stable training of RBMs.\n" + "abstract": "Abstract:Boltzmann machines are often used as building blocks in greedy learning of deep networks. However, training even a simplified model, known as restricted Boltzmann machine (RBM), can be extremely laborious: Traditional learning algorithms often converge only with the right choice of the learning rate scheduling and the scale of the initial weights. They are also sensitive to specific data representation: An equivalent RBM can be obtained by flipping some bits and changing the weights and biases accordingly, but traditional learning rules are not invariant to such transformations. Without careful tuning of these training settings, traditional algorithms can easily get stuck at plateaus or even diverge. In this work, we present an enhanced gradient which is derived such that it is invariant to bit-flipping transformations. We also propose a way to automatically adjust the learning rate by maximizing a local likelihood estimate. Our experiments confirm that the proposed improvements yield more stable training of RBMs." }, { "author": [ @@ -229,14 +229,14 @@ "Kolmogorov, Vladimir" ], "paper_title": "Dynamic Tree Block Coordinate Ascent", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Neural Computing and Applications", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 113, + "page_form": 113, "page_to": 120, "totle_page": 8, "language": "en", - "abstract": "Abstract:This paper proposes a novel Linear Programming (LP) based algorithm, called Dynamic Tree-Block Coordinate Ascent (DTBCA), for perfroming maximum a posteriori (MAP) inference in probabilistic graphical models. Unlike traditional message passing algorithms, which operate unifromly on the whole factor graph, our method dynamically chooses regions of the factor graph on which to focus message-passing efforts. We propose two criteria for selecting regions, including an efficiently computable upper-bound on the increase in the objective possible by passing messages in any particular region. This bound is derived from the theory of primal-dual methods from combinatorial optimization, and the forest that maximizes the bounds can be chosen efficiently using a maximum-spanning-tree-like algorithm. Experimental results show that our dynamic schedules significantly speed up state-of-the- art LP-based message-passing algorithms on a wide variety of real-world problems.\n" + "abstract": "Abstract:This paper proposes a novel Linear Programming (LP) based algorithm, called Dynamic Tree-Block Coordinate Ascent (DTBCA), for performing maximum a posteriori (MAP) inference in probabilistic graphical models. Unlike traditional message passing algorithms, which operate uniformly on the whole factor graph, our method dynamically chooses regions of the factor graph on which to focus message-passing efforts. We propose two criteria for selecting regions, including an efficiently computable upper-bound on the increase in the objective possible by passing messages in any particular region. This bound is derived from the theory of primal-dual methods from combinatorial optimization, and the forest that maximizes the bounds can be chosen efficiently using a maximum-spanning-tree-like algorithm. Experimental results show that our dynamic schedules significantly speed up state-of-the- art LP-based message-passing algorithms on a wide variety of real-world problems." }, { "author": [ @@ -244,14 +244,14 @@ "Orecchia, Lorenzo" ], "paper_title": "Implementing regularization implicitly via approximate eigenvector computation", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "European Journal of Information Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 121, + "page_form": 121, "page_to": 128, "totle_page": 8, "language": "en", - "abstract": "Abstract:Regularization is a powerful technique for extracting useful infromation from noisy data. Typically, it is implemented by adding some sort of norm constraint to an objective function and then exactly optimizing the modified objective function. This procedure often leads to optimization problems that are computationally more expensive than the original problem, a fact that is clearly problematic if one is interested in large-scale applications. On the other hand, a large body of empirical work has demonstrated that heuristics, and in some cases approximation algorithms, developed to speed up computations sometimes have the side-effect of perfroming regularization implicitly. Thus, we consider the question: What is the regularized optimization objective that an approximation algorithm is exactly optimizing? We address this question in the context of computing approximations to the smallest nontrivial eigenvector of a graph Laplacian; and we consider three random-walk-based procedures: one based on the heat kernel of the graph, one based on computing the the PageRank vector associated with the graph, and one based on a truncated lazy random walk. In each case, we provide a precise characterization of the manner in which the approximation method can be viewed as implicitly computing the exact solution to a regularized problem. Interestingly, the regularization is not on the usual vector from of the optimization problem, but instead it is on a related semidefinite program.\n" + "abstract": "Abstract:Regularization is a powerful technique for extracting useful information from noisy data. Typically, it is implemented by adding some sort of norm constraint to an objective function and then exactly optimizing the modified objective function. This procedure often leads to optimization problems that are computationally more expensive than the original problem, a fact that is clearly problematic if one is interested in large-scale applications. On the other hand, a large body of empirical work has demonstrated that heuristics, and in some cases approximation algorithms, developed to speed up computations sometimes have the side-effect of performing regularization implicitly. Thus, we consider the question: What is the regularized optimization objective that an approximation algorithm is exactly optimizing? We address this question in the context of computing approximations to the smallest nontrivial eigenvector of a graph Laplacian; and we consider three random-walk-based procedures: one based on the heat kernel of the graph, one based on computing the the PageRank vector associated with the graph, and one based on a truncated lazy random walk. In each case, we provide a precise characterization of the manner in which the approximation method can be viewed as implicitly computing the exact solution to a regularized problem. Interestingly, the regularization is not on the usual vector form of the optimization problem, but instead it is on a related semidefinite program." }, { "author": [ @@ -261,14 +261,14 @@ "Manning, Chris" ], "paper_title": "Parsing Natural Scenes and Natural Language with Recursive Neural Networks", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Internet Mathematics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 129, + "page_form": 129, "page_to": 136, "totle_page": 8, "language": "en", - "abstract": "Abstract:Recursive structure is commonly found in the inputs of different modalities such as natural scene images or natural language sentences. Discovering this recursive structure helps us to not only identify the units that an image or sentence contains but also how they interact to from a whole. We introduce a max-margin structure prediction architecture based on recursive neural networks that can successfully recover such structure both in complex scene images as well as sentences. The same algorithm can be used both to provide a competitive syntactic parser for natural language sentences from the Penn Treebank and to outperfrom alternative approaches for semantic scene segmentation, annotation and classification. For segmentation and annotation our algorithm obtains a new level of state-of-the-art perfromance on the Stanford background dataset (78.1%). The features from the image parse tree outperfrom Gist descriptors for scene classification by 4%.\n" + "abstract": "Abstract:Recursive structure is commonly found in the inputs of different modalities such as natural scene images or natural language sentences. Discovering this recursive structure helps us to not only identify the units that an image or sentence contains but also how they interact to form a whole. We introduce a max-margin structure prediction architecture based on recursive neural networks that can successfully recover such structure both in complex scene images as well as sentences. The same algorithm can be used both to provide a competitive syntactic parser for natural language sentences from the Penn Treebank and to outperform alternative approaches for semantic scene segmentation, annotation and classification. For segmentation and annotation our algorithm obtains a new level of state-of-the-art performance on the Stanford background dataset (78.1%). The features from the image parse tree outperform Gist descriptors for scene classification by 4%." }, { "author": [ @@ -276,14 +276,14 @@ "Barto, Andrew" ], "paper_title": "Conjugate Markov Decision Processes ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Transactions on Rough Sets", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 137, + "page_form": 137, "page_to": 144, "totle_page": 8, "language": "en", - "abstract": "Abstract:Many open problems involve the search for a mapping that is used by an algorithm solving an MDP. Useful mappings are often from the state set to some other set. Examples include representation discovery (a mapping to a feature space) and skill discovery (a mapping to skill termination probabilities). Different mappings result in algorithms achieving varying expected returns. In this paper we present a novel approach to the search for any mapping used by any algorithm attempting to solve an MDP, for that which results in maximum expected return.\n" + "abstract": "Abstract:Many open problems involve the search for a mapping that is used by an algorithm solving an MDP. Useful mappings are often from the state set to some other set. Examples include representation discovery (a mapping to a feature space) and skill discovery (a mapping to skill termination probabilities). Different mappings result in algorithms achieving varying expected returns. In this paper we present a novel approach to the search for any mapping used by any algorithm attempting to solve an MDP, for that which results in maximum expected return." }, { "author": [ @@ -291,28 +291,28 @@ "Boutilier, Craig" ], "paper_title": "Learning Mallows Models with Pairwise Preferences", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IMPACT Impact of Computing in Science and Engineering", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 145, + "page_form": 145, "page_to": 152, "totle_page": 8, "language": "en", - "abstract": "Abstract:Learning preference distributions is a key problem in many areas (e.g., recommender systems, IR, social choice). However, existing methods require restrictive data models for evidence about user preferences. We relax these restrictions by considering as data arbitrary pairwise comparisons---the fundamental building blocks of ordinal rankings. We develop the first algorithms for learning Mallows models (and mixtures) with pairwise comparisons. At the heart is a new algorithm, the generalized repeated insertion model (GRIM), for sampling from arbitrary ranking distributions. We develop approximate samplers that are exact for many important special cases---and have provable bounds with pairwise evidence---and derive algorithms for evaluating log-likelihood, learning Mallows mixtures, and non-parametric estimation. Experiments on large, real-world datasets show the effectiveness of our approach.\n" + "abstract": "Abstract:Learning preference distributions is a key problem in many areas (e.g., recommender systems, IR, social choice). However, existing methods require restrictive data models for evidence about user preferences. We relax these restrictions by considering as data arbitrary pairwise comparisons---the fundamental building blocks of ordinal rankings. We develop the first algorithms for learning Mallows models (and mixtures) with pairwise comparisons. At the heart is a new algorithm, the generalized repeated insertion model (GRIM), for sampling from arbitrary ranking distributions. We develop approximate samplers that are exact for many important special cases---and have provable bounds with pairwise evidence---and derive algorithms for evaluating log-likelihood, learning Mallows mixtures, and non-parametric estimation. Experiments on large, real-world datasets show the effectiveness of our approach." }, { "author": [ "Scott, Clayton" ], "paper_title": "Surrogate losses and regret bounds for cost-sensitive classification with example-dependent costs ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Cognition, Technology & Work", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 153, + "page_form": 153, "page_to": 160, "totle_page": 8, "language": "en", - "abstract": "Abstract:We study surrogate losses in the context of cost-sensitive classification with example-dependent costs, a problem also known as regression level set estimation. We give sufficient conditions on the surrogate loss for the existence of a surrogate regret bound. Such bounds imply that as the surrogate risk tends to its optimal value, so too does the expected misclassification cost. Our sufficient conditions encompass example-dependent versions of the hinge, exponential, and other common losses. These results provide theoretical justification for some previously proposed surrogate-based algorithms, and suggests others that have not yet been developed.\n" + "abstract": "Abstract:We study surrogate losses in the context of cost-sensitive classification with example-dependent costs, a problem also known as regression level set estimation. We give sufficient conditions on the surrogate loss for the existence of a surrogate regret bound. Such bounds imply that as the surrogate risk tends to its optimal value, so too does the expected misclassification cost. Our sufficient conditions encompass example-dependent versions of the hinge, exponential, and other common losses. These results provide theoretical justification for some previously proposed surrogate-based algorithms, and suggests others that have not yet been developed." }, { "author": [ @@ -321,14 +321,14 @@ "Ramakrishnan, Ganesh" ], "paper_title": "Efficient Rule Ensemble Learning using Hierarchical Kernels ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal of Information Technology and Decision Making", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 161, + "page_form": 161, "page_to": 168, "totle_page": 8, "language": "en", - "abstract": "Abstract:This paper addresses the problem of Rule Ensemble Learning (REL), where the goal is simultaneous discovery of a small set of simple rules and their optimal weights that lead to good generalization. Rules are assumed to be conjunctions of basic propositions concerning the values taken by the input features. From the perspectives of interpretability as well as generalization, it is highly desirable to construct rule ensembles with low training error, having rules that are i) simple, {\\em i.e.}, involve few conjunctions and ii) few in number. We propose to explore the (exponentially) large feature space of all possible conjunctions optimally and efficiently by employing the recently introduced Hierarchical Kernel Learning (HKL) framework. The regularizer employed in the HKL fromulation can be interpreted as a potential for discouraging selection of rules involving large number of conjunctions -- justifying its suitability for constructing rule ensembles. Simulation results show that the proposed approach improves over state-of-the-art REL algorithms in terms of generalization and indeed learns simple rules. Although this is encouraging, it can be shown that HKL selects a conjunction only if all its subsets are selected, and this is highly undesirable. We propose a novel convex fromulation which alleviates this problem and generalizes the HKL framework. The main technical contribution of this paper is an efficient mirror-descent based active set algorithm for solving the new fromulation. Empirical evaluations on REL problems illustrate the utility of generalized HKL.\n" + "abstract": "Abstract:This paper addresses the problem of Rule Ensemble Learning (REL), where the goal is simultaneous discovery of a small set of simple rules and their optimal weights that lead to good generalization. Rules are assumed to be conjunctions of basic propositions concerning the values taken by the input features. From the perspectives of interpretability as well as generalization, it is highly desirable to construct rule ensembles with low training error, having rules that are i) simple, {\\em i.e.}, involve few conjunctions and ii) few in number. We propose to explore the (exponentially) large feature space of all possible conjunctions optimally and efficiently by employing the recently introduced Hierarchical Kernel Learning (HKL) framework. The regularizer employed in the HKL formulation can be interpreted as a potential for discouraging selection of rules involving large number of conjunctions -- justifying its suitability for constructing rule ensembles. Simulation results show that the proposed approach improves over state-of-the-art REL algorithms in terms of generalization and indeed learns simple rules. Although this is encouraging, it can be shown that HKL selects a conjunction only if all its subsets are selected, and this is highly undesirable. We propose a novel convex formulation which alleviates this problem and generalizes the HKL framework. The main technical contribution of this paper is an efficient mirror-descent based active set algorithm for solving the new formulation. Empirical evaluations on REL problems illustrate the utility of generalized HKL." }, { "author": [ @@ -339,14 +339,14 @@ "Xing, Eric" ], "paper_title": "An Augmented Lagrangian Approach to Constrained MAP Inference ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal of Knowledge and Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 169, + "page_form": 169, "page_to": 176, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose a new fast algorithm for approximate MAP inference on factor graphs, which combines augmented Lagrangian optimization with the dual decomposition method. Each slave subproblem is given a quadratic penalty, which pushes toward faster consensus than in previous subgradient approaches. Our algorithm is provably convergent, parallelizable, and suitable for fine decompositions of the graph. We show how it can efficiently handle problems with (possibly global) structural constraints via simple sort operations. Experiments on synthetic and real-world data show that our approach compares favorably with the state-of-the-art.\n" + "abstract": "Abstract:We propose a new fast algorithm for approximate MAP inference on factor graphs, which combines augmented Lagrangian optimization with the dual decomposition method. Each slave subproblem is given a quadratic penalty, which pushes toward faster consensus than in previous subgradient approaches. Our algorithm is provably convergent, parallelizable, and suitable for fine decompositions of the graph. We show how it can efficiently handle problems with (possibly global) structural constraints via simple sort operations. Experiments on synthetic and real-world data show that our approach compares favorably with the state-of-the-art." }, { "author": [ @@ -354,14 +354,14 @@ "Tsitsiklis, John" ], "paper_title": "Mean-Variance Optimization in Markov Decision Processes ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal of Web and Grid Services", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 177, + "page_form": 177, "page_to": 184, "totle_page": 8, "language": "en", - "abstract": "Abstract:We consider finite horizon Markov decision processes under perfromance measures that involve both the mean and the variance of the cumulative reward. We show that either randomized or history-based policies can improve perfromance. We prove that the complexity of computing a policy that maximizes the mean reward under a variance constraint is NP-hard for some cases, and strongly NP-hard for others. We finally offer pseudo-polynomial exact and approximation algorithms.\n" + "abstract": "Abstract:We consider finite horizon Markov decision processes under performance measures that involve both the mean and the variance of the cumulative reward. We show that either randomized or history-based policies can improve performance. We prove that the complexity of computing a policy that maximizes the mean reward under a variance constraint is NP-hard for some cases, and strongly NP-hard for others. We finally offer pseudo-polynomial exact and approximation algorithms." }, { "author": [ @@ -369,42 +369,42 @@ "Prakash, B. Aditya" ], "paper_title": "Time Series Clustering: Complex is Simpler!", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Robotica", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 185, + "page_form": 185, "page_to": 192, "totle_page": 8, "language": "en", - "abstract": "Abstract:Given a motion capture sequence, how to identify the category of the motion? Classifying human motions is a critical task in motion editing and synthesizing, for which manual labeling is clearly inefficient for large databases. Here we study the general problem of time series clustering. We propose a novel method of clustering time series that can (a) learn joint temporal dynamics in the data; (b) handle time lags; and (c) produce interpretable features. We achieve this by developing complex-valued linear dynamical systems (CLDS), which include real-valued Kalman filters as a special case; our advantage is that the transition matrix is simpler (just diagonal), and the transmission one easier to interpret. We then present Complex-Fit, a novel EM algorithm to learn the parameters for the general model and its special case for clustering. Our approach produces significant improvement in clustering quality, 1.5 to 5 times better than well-known competitors on real motion capture sequences.\n" + "abstract": "Abstract:Given a motion capture sequence, how to identify the category of the motion? Classifying human motions is a critical task in motion editing and synthesizing, for which manual labeling is clearly inefficient for large databases. Here we study the general problem of time series clustering. We propose a novel method of clustering time series that can (a) learn joint temporal dynamics in the data; (b) handle time lags; and (c) produce interpretable features. We achieve this by developing complex-valued linear dynamical systems (CLDS), which include real-valued Kalman filters as a special case; our advantage is that the transition matrix is simpler (just diagonal), and the transmission one easier to interpret. We then present Complex-Fit, a novel EM algorithm to learn the parameters for the general model and its special case for clustering. Our approach produces significant improvement in clustering quality, 1.5 to 5 times better than well-known competitors on real motion capture sequences." }, { "author": [ "Gould, Stephen" ], "paper_title": "Max-margin Learning for Lower Linear Envelope Potentials in Binary Markov Random Fields", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Machine Intelligence and Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 193, + "page_form": 193, "page_to": 200, "totle_page": 8, "language": "en", - "abstract": "Abstract:The standard approach to max-margin parameter learning for Markov random fields (MRFs) involves incrementally adding the most violated constraints during each iteration of the algorithm. This requires exact MAP inference, which is intractable for many classes of MRF. In this paper, we propose an exact MAP inference algorithm for binary MRFs containing a class of higher-order models, known as lower linear envelope potentials. Our algorithm is polynomial in the number of variables and number of linear envelope functions. With tractable inference in hand, we show how the parameters and corresponding feature vectors can be represented in a max-margin framework for efficiently learning lower linear envelope potentials.\n" + "abstract": "Abstract:The standard approach to max-margin parameter learning for Markov random fields (MRFs) involves incrementally adding the most violated constraints during each iteration of the algorithm. This requires exact MAP inference, which is intractable for many classes of MRF. In this paper, we propose an exact MAP inference algorithm for binary MRFs containing a class of higher-order models, known as lower linear envelope potentials. Our algorithm is polynomial in the number of variables and number of linear envelope functions. With tractable inference in hand, we show how the parameters and corresponding feature vectors can be represented in a max-margin framework for efficiently learning lower linear envelope potentials." }, { "author": [ "Clark, Alexander" ], "paper_title": "Inference of Inversion Transduction Grammars ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IEEE Transactions on Aerospace and Navigational Electronics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 201, + "page_form": 201, "page_to": 208, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present the first polynomial algorithm for learning a class of inversion transduction grammars (ITGs) that implement context free transducers -- functions from strings to strings. The class of transductions that we can learn properly includes all subsequential transductions. These algorithms are based on a generalisation of distributional learning; we prove correctness of our algorithm under an identification in the limit model.\n" + "abstract": "Abstract:We present the first polynomial algorithm for learning a class of inversion transduction grammars (ITGs) that implement context free transducers -- functions from strings to strings. The class of transductions that we can learn properly includes all subsequential transductions. These algorithms are based on a generalisation of distributional learning; we prove correctness of our algorithm under an identification in the limit model." }, { "author": [ @@ -413,28 +413,28 @@ "Chen, SongCan" ], "paper_title": "BCDNPKL: Scalable Non-Parametric Kernel Learning Using Block Coordinate Descent", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Transactions on Machine Learning and Data Mining", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 209, + "page_form": 209, "page_to": 216, "totle_page": 8, "language": "en", - "abstract": "Abstract:Most existing approaches for non-parametric kernel learning (NPKL) suffer from expensive computation, which would limit their applications to large-scale problems. To address the scalability problem of NPKL, we propose a novel algorithm called BCDNPKL, which is very efficient and scalable. Superior to most existing approaches, BCDNPKL keeps away from semidefinite programming (SDP) and eigen-decomposition, which benefits from two findings: 1) The original SDP framework of NPKL can be reduced into a far smaller-sized counterpart which is corresponding to the sub-kernel (referred to as boundary kernel) learning; 2) The sub-kernel learning can be efficiently solved by using the proposed block coordinate descent (BCD) technique. We provide a fromal proof of global convergence for the proposed BCDNPKL algorithm. The extensive experiments verify the scalability and effectiveness of BCDNPKL, compared with the state-of-the-art algorithms.\n" + "abstract": "Abstract:Most existing approaches for non-parametric kernel learning (NPKL) suffer from expensive computation, which would limit their applications to large-scale problems. To address the scalability problem of NPKL, we propose a novel algorithm called BCDNPKL, which is very efficient and scalable. Superior to most existing approaches, BCDNPKL keeps away from semidefinite programming (SDP) and eigen-decomposition, which benefits from two findings: 1) The original SDP framework of NPKL can be reduced into a far smaller-sized counterpart which is corresponding to the sub-kernel (referred to as boundary kernel) learning; 2) The sub-kernel learning can be efficiently solved by using the proposed block coordinate descent (BCD) technique. We provide a formal proof of global convergence for the proposed BCDNPKL algorithm. The extensive experiments verify the scalability and effectiveness of BCDNPKL, compared with the state-of-the-art algorithms." }, { "author": [ "der Maaten, Laurens Van" ], "paper_title": "Learning Discriminative Fisher Kernels", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Neural Information Processing Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 217, + "page_form": 217, "page_to": 224, "totle_page": 8, "language": "en", - "abstract": "Abstract:Fisher kernels provide a commonly used vectorial representation of structured objects. The paper presents a technique that exploits label infromation to improve the object representation of Fisher kernels by employing ideas from metric learning. In particular, the new technique trains a generative model in such a way that the distance between the log-likelihood gradients induced by two objects with the same label is as small as possible, and the distance between the gradients induced by two objects with different labels is as large as possible. We illustrate the strong perfromance of classifiers trained on the resulting object representations on problems in handwriting recognition, speech recognition, facial expression analysis, and bio-infromatics.\n" + "abstract": "Abstract:Fisher kernels provide a commonly used vectorial representation of structured objects. The paper presents a technique that exploits label information to improve the object representation of Fisher kernels by employing ideas from metric learning. In particular, the new technique trains a generative model in such a way that the distance between the log-likelihood gradients induced by two objects with the same label is as small as possible, and the distance between the gradients induced by two objects with different labels is as large as possible. We illustrate the strong performance of classifiers trained on the resulting object representations on problems in handwriting recognition, speech recognition, facial expression analysis, and bio-informatics." }, { "author": [ @@ -442,14 +442,14 @@ "von Luxburg, Ulrike" ], "paper_title": "Pruning nearest neighbor cluster trees ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Machine Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 225, + "page_form": 225, "page_to": 232, "totle_page": 8, "language": "en", - "abstract": "Abstract:Nearest neighbor ($k$-NN) graphs are widely used in machine learning and data mining applications, and our aim is to better understand what they reveal about the cluster structure of the unknown underlying distribution of points. Moreover, is it possible to identify spurious structures that might arise due to sampling variability? Our first contribution is a statistical analysis that reveals how certain subgraphs of a $k$-NN graph from a consistent estimator of the cluster tree of the underlying distribution of points. Our second and perhaps most important contribution is the following finite sample guarantee. We carefully work out the tradeoff between aggressive and conservative pruning and are able to guarantee the removal of all spurious cluster structures while at the same time guaranteeing the recovery of salient clusters. This is the first such finite sample result in the context of clustering.\n" + "abstract": "Abstract:Nearest neighbor ($k$-NN) graphs are widely used in machine learning and data mining applications, and our aim is to better understand what they reveal about the cluster structure of the unknown underlying distribution of points. Moreover, is it possible to identify spurious structures that might arise due to sampling variability? Our first contribution is a statistical analysis that reveals how certain subgraphs of a $k$-NN graph form a consistent estimator of the cluster tree of the underlying distribution of points. Our second and perhaps most important contribution is the following finite sample guarantee. We carefully work out the tradeoff between aggressive and conservative pruning and are able to guarantee the removal of all spurious cluster structures while at the same time guaranteeing the recovery of salient clusters. This is the first such finite sample result in the context of clustering." }, { "author": [ @@ -459,14 +459,14 @@ "Yang, Tianbao" ], "paper_title": "Online AUC Maximization ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Uncertainty in Artificial Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 233, + "page_form": 233, "page_to": 240, "totle_page": 8, "language": "en", - "abstract": "Abstract:Most studies of online learning measure the perfromance of a learner by classification accuracy, which is inappropriate for applications where the data are unevenly distributed among different classes. We address this limitation by developing online learning algorithm for maximizing Area Under the ROC curve (AUC), a metric that is widely used for measuring the classification perfromance for imbalanced data distributions. The key challenge of online AUC maximization is that it needs to optimize the pairwise loss between two instances from different classes. This is in contrast to the classical setup of online learning where the overall loss is a sum of losses over individual training examples. We address this challenge by exploiting the reservoir sampling technique, and present two algorithms for online AUC maximization with theoretic perfromance guarantee. Extensive experimental studies confirm the effectiveness and the efficiency of the proposed algorithms for maximizing AUC.\n" + "abstract": "Abstract:Most studies of online learning measure the performance of a learner by classification accuracy, which is inappropriate for applications where the data are unevenly distributed among different classes. We address this limitation by developing online learning algorithm for maximizing Area Under the ROC curve (AUC), a metric that is widely used for measuring the classification performance for imbalanced data distributions. The key challenge of online AUC maximization is that it needs to optimize the pairwise loss between two instances from different classes. This is in contrast to the classical setup of online learning where the overall loss is a sum of losses over individual training examples. We address this challenge by exploiting the reservoir sampling technique, and present two algorithms for online AUC maximization with theoretic performance guarantee. Extensive experimental studies confirm the effectiveness and the efficiency of the proposed algorithms for maximizing AUC." }, { "author": [ @@ -474,14 +474,14 @@ "Joachims, Thorsten" ], "paper_title": "Beat the Mean Bandit ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Intelligent RObots and IROS", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 241, + "page_form": 241, "page_to": 248, "totle_page": 8, "language": "en", - "abstract": "Abstract:The Dueling Bandits Problem is an online learning framework in which actions are restricted to noisy comparisons between pairs of strategies (also known as bandits). It models settings where absolute rewards are difficult to elicit but pairwise preferences are readily available. In this paper, we extend the Dueling Bandits Problem to a relaxed setting where preference magnitudes can violate transitivity. We present the first algorithm for this more general Dueling Bandits Problem and provide theoretical guarantees in both the online and the PAC settings. Furthermore, we show that the new algorithm has stronger guarantees than existing results even in the original Dueling Bandits Problem, which we validate empirically.\n" + "abstract": "Abstract:The Dueling Bandits Problem is an online learning framework in which actions are restricted to noisy comparisons between pairs of strategies (also known as bandits). It models settings where absolute rewards are difficult to elicit but pairwise preferences are readily available. In this paper, we extend the Dueling Bandits Problem to a relaxed setting where preference magnitudes can violate transitivity. We present the first algorithm for this more general Dueling Bandits Problem and provide theoretical guarantees in both the online and the PAC settings. Furthermore, we show that the new algorithm has stronger guarantees than existing results even in the original Dueling Bandits Problem, which we validate empirically." }, { "author": [ @@ -489,28 +489,28 @@ "Jie, Luo" ], "paper_title": "Ultra-Fast Optimization Algorithm for Sparse Multi Kernel Learning", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 249, + "page_form": 249, "page_to": 256, "totle_page": 8, "language": "en", - "abstract": "Abstract:Many state-of-the-art approaches for Multi Kernel Learning (MKL) struggle at finding a compromise between perfromance, sparsity of the solution and speed of the optimization process. In this paper we look at the MKL problem at the same time from a learning and optimization point of view. So, instead of designing a regularizer and then struggling to find an efficient method to minimize it, we design the regularizer while keeping the optimization algorithm in mind. Hence, we introduce a novel MKL fromulation, which mixes elements of p-norm and elastic-net kind of regularization. We also propose a fast stochastic gradient descent method that solves the novel MKL fromulation. We show theoretically and empirically that our method has 1) state-of-the-art perfromance on many classification tasks; 2) exact sparse solutions with a tunable level of sparsity; 3) a convergence rate bound that depends only logarithmically on the number of kernels used, and is independent of the sparsity required; 4) independence on the particular convex loss function used.\n" + "abstract": "Abstract:Many state-of-the-art approaches for Multi Kernel Learning (MKL) struggle at finding a compromise between performance, sparsity of the solution and speed of the optimization process. In this paper we look at the MKL problem at the same time from a learning and optimization point of view. So, instead of designing a regularizer and then struggling to find an efficient method to minimize it, we design the regularizer while keeping the optimization algorithm in mind. Hence, we introduce a novel MKL formulation, which mixes elements of p-norm and elastic-net kind of regularization. We also propose a fast stochastic gradient descent method that solves the novel MKL formulation. We show theoretically and empirically that our method has 1) state-of-the-art performance on many classification tasks; 2) exact sparse solutions with a tunable level of sparsity; 3) a convergence rate bound that depends only logarithmically on the number of kernels used, and is independent of the sparsity required; 4) independence on the particular convex loss function used." }, { "author": [ "Potetz, Brian" ], "paper_title": "Estimating the Bayes Point Using Linear Knapsack Problems ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Symposium on Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 257, + "page_form": 257, "page_to": 264, "totle_page": 8, "language": "en", - "abstract": "Abstract:A Bayes Point machine is a binary classifier that approximates the Bayes-optimal classifier by estimating the mean of the posterior distribution of classifier parameters. Past Bayes Point machines have overcome the intractability of this goal by using message passing techniques that approximate the posterior of the classifier parameters as a Gaussian distribution. In this paper, we investigate alternative message passing approaches that do not rely on Gaussian approximation. To make this possible, we introduce a new computational shortcut based on linear multiple-choice knapsack problems that reduces the complexity of computing Bayes Point belief propagation messages from exponential to linear in the number of data features. Empirical tests of our approach show significant improvement in linear classification over both soft-margin SVMs and Expectation Propagation Bayes Point machines for several real-world UCI datasets.\n" + "abstract": "Abstract:A Bayes Point machine is a binary classifier that approximates the Bayes-optimal classifier by estimating the mean of the posterior distribution of classifier parameters. Past Bayes Point machines have overcome the intractability of this goal by using message passing techniques that approximate the posterior of the classifier parameters as a Gaussian distribution. In this paper, we investigate alternative message passing approaches that do not rely on Gaussian approximation. To make this possible, we introduce a new computational shortcut based on linear multiple-choice knapsack problems that reduces the complexity of computing Bayes Point belief propagation messages from exponential to linear in the number of data features. Empirical tests of our approach show significant improvement in linear classification over both soft-margin SVMs and Expectation Propagation Bayes Point machines for several real-world UCI datasets." }, { "author": [ @@ -522,14 +522,14 @@ "Ng, Andrew" ], "paper_title": "On optimization methods for deep learning ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Computational Learning Theory", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 265, + "page_form": 265, "page_to": 272, "totle_page": 8, "language": "en", - "abstract": "Abstract:The predominant methodology in training deep learning advocates the use of stochastic gradient descent methods (SGDs). Despite its ease of implementation, SGDs are difficult to tune and parallelize. These problems make it challenging to develop, debug and scale up deep learning algorithms with SGDs. In this paper, we show that more sophisticated off-the-shelf optimization methods such as Limited memory BFGS (L-BFGS) and Conjugate gradient (CG) with linesearch can significantly simplify and speed up the process of pretraining deep algorithms. In our experiments, the difference between L-BFGS/CG and SGDs are more pronounced if we consider algorithmic extensions (e.g., sparsity regularization) and hardware extensions (e.g., GPUs or computer clusters). Our experiments with distributed optimization support the use of L-BFGS with locally connected networks and convolutional neural networks. Using L-BFGS, our convolutional network model achieves 0.69\\% on the standard MNIST dataset. This is a state-of-the-art result on MNIST among algorithms that do not use distortions or pretraining.\n" + "abstract": "Abstract:The predominant methodology in training deep learning advocates the use of stochastic gradient descent methods (SGDs). Despite its ease of implementation, SGDs are difficult to tune and parallelize. These problems make it challenging to develop, debug and scale up deep learning algorithms with SGDs. In this paper, we show that more sophisticated off-the-shelf optimization methods such as Limited memory BFGS (L-BFGS) and Conjugate gradient (CG) with linesearch can significantly simplify and speed up the process of pretraining deep algorithms. In our experiments, the difference between L-BFGS/CG and SGDs are more pronounced if we consider algorithmic extensions (e.g., sparsity regularization) and hardware extensions (e.g., GPUs or computer clusters). Our experiments with distributed optimization support the use of L-BFGS with locally connected networks and convolutional neural networks. Using L-BFGS, our convolutional network model achieves 0.69\\% on the standard MNIST dataset. This is a state-of-the-art result on MNIST among algorithms that do not use distortions or pretraining." }, { "author": [ @@ -537,14 +537,14 @@ "Gentile, Claudio" ], "paper_title": "Multiclass Classification with Bandit Feedback using Adaptive Regularization ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 273, + "page_form": 273, "page_to": 280, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present a new multiclass algorithm in the bandit framework, where after making a prediction, the learning algorithm receives only partial feedback, i.e., a single bit of right-or-wrong, rather then the true label. Our algorithm is based on the second-order Perceptron, and uses upper-confidence bounds to trade off exploration and exploitation. We analyze this algorithm in a partial adversarial setting, where instances are chosen adversarially, while the labels are chosen according to a linear probabilistic model, which is also chosen adversarially. From the theoretical viewpoint, we show a regret of O(\\sqrt{T}\\log(T)), which improves over the current best bounds of O(T^{2/3}) in the fully adversarial setting. We evaluate our algorithm on nine real-world text classification problems, obtaining state-of-the-art results, even compared with non-bandit online algorithms, especially when label noise is introduced.\n" + "abstract": "Abstract:We present a new multiclass algorithm in the bandit framework, where after making a prediction, the learning algorithm receives only partial feedback, i.e., a single bit of right-or-wrong, rather then the true label. Our algorithm is based on the second-order Perceptron, and uses upper-confidence bounds to trade off exploration and exploitation. We analyze this algorithm in a partial adversarial setting, where instances are chosen adversarially, while the labels are chosen according to a linear probabilistic model, which is also chosen adversarially. From the theoretical viewpoint, we show a regret of O(\\sqrt{T}\\log(T)), which improves over the current best bounds of O(T^{2/3}) in the fully adversarial setting. We evaluate our algorithm on nine real-world text classification problems, obtaining state-of-the-art results, even compared with non-bandit online algorithms, especially when label noise is introduced." }, { "author": [ @@ -552,14 +552,14 @@ "Long, Phil" ], "paper_title": "On the Necessity of Irrelevant Variables ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Document Analysis and Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 281, + "page_form": 281, "page_to": 288, "totle_page": 8, "language": "en", - "abstract": "Abstract:This work explores the effects of relevant and irrelevant boolean variables on the accuracy of classifiers. The analysis uses the assumption that the variables are conditionally independent given the class, and focuses on a natural family of learning algorithms for such sources when the relevant variables have a small advantage over random guessing. The main result is that algorithms relying predominately on irrelevant variables have error probabilities that quickly go to 0 in situations where algorithms that limit the use of irrelevant variables have errors bounded below by a positive constant. We also show that accurate learning is possible even when there are so few examples that one cannot determine with high confidence whether or not any individual variable is relevant.\n" + "abstract": "Abstract:This work explores the effects of relevant and irrelevant boolean variables on the accuracy of classifiers. The analysis uses the assumption that the variables are conditionally independent given the class, and focuses on a natural family of learning algorithms for such sources when the relevant variables have a small advantage over random guessing. The main result is that algorithms relying predominately on irrelevant variables have error probabilities that quickly go to 0 in situations where algorithms that limit the use of irrelevant variables have errors bounded below by a positive constant. We also show that accurate learning is possible even when there are so few examples that one cannot determine with high confidence whether or not any individual variable is relevant." }, { "author": [ @@ -567,14 +567,14 @@ "Chopin, Nicolas" ], "paper_title": "ABC-EP: Expectation Propagation for Likelihood-free Bayesian Computation ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Int. Conference on Artificial Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 289, + "page_form": 289, "page_to": 296, "totle_page": 8, "language": "en", - "abstract": "Abstract:Many statistical models of interest to the natural and social sciences have no tractable likelihood function. Until recently, Bayesian inference for such models was thought infeasible. Pritchard et al. (1999) introduced an algorithm known as ABC, for Approximate Bayesian Computation, that enables Bayesian computation in such models. Despite steady progress since this first breakthrough, such as the adaptation of MCMC and Sequential Monte Carlo techniques to likelihood-free inference, state-of-the art methods remain notoriously hard to use and require enormous computation times. Among other issues, one faces the difficult task of finding appropriate summary statistics for the model, and tuning the algorithm can be time-consuming when little prior infromation is available. We show that Expectation Propagation, a widely successful approximate inference technique, can be adapted to the likelihood-free context. The resulting algorithm does not require summary statistics, is an order of magnitude faster than existing techniques, and remains usable when prior infromation is vague.\n" + "abstract": "Abstract:Many statistical models of interest to the natural and social sciences have no tractable likelihood function. Until recently, Bayesian inference for such models was thought infeasible. Pritchard et al. (1999) introduced an algorithm known as ABC, for Approximate Bayesian Computation, that enables Bayesian computation in such models. Despite steady progress since this first breakthrough, such as the adaptation of MCMC and Sequential Monte Carlo techniques to likelihood-free inference, state-of-the art methods remain notoriously hard to use and require enormous computation times. Among other issues, one faces the difficult task of finding appropriate summary statistics for the model, and tuning the algorithm can be time-consuming when little prior information is available. We show that Expectation Propagation, a widely successful approximate inference technique, can be adapted to the likelihood-free context. The resulting algorithm does not require summary statistics, is an order of magnitude faster than existing techniques, and remains usable when prior information is vague." }, { "author": [ @@ -585,14 +585,14 @@ "Shanian, Sara" ], "paper_title": "A PAC-Bayes Sample-compression Approach to Kernel Methods ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Independent Component Analysis", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 297, + "page_form": 297, "page_to": 304, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose a PAC-Bayes sample compression approach to kernel methods that can accommodate any bounded similarity function and show that the support vector machine (SVM) classifier is a particular case of a more general class of data-dependent classifiers known as majority votes of sample-compressed classifiers. We provide novel risk bounds for these majority votes and learning algorithms that minimize these bounds.\n" + "abstract": "Abstract:We propose a PAC-Bayes sample compression approach to kernel methods that can accommodate any bounded similarity function and show that the support vector machine (SVM) classifier is a particular case of a more general class of data-dependent classifiers known as majority votes of sample-compressed classifiers. We provide novel risk bounds for these majority votes and learning algorithms that minimize these bounds." }, { "author": [ @@ -601,14 +601,14 @@ "Meir, Ron" ], "paper_title": "Integrating Partial Model Knowledge in Model Free RL Algorithms", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "DAGM Symposium Symposium for Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 305, + "page_form": 305, "page_to": 312, "totle_page": 8, "language": "en", - "abstract": "Abstract:In reinforcement learning an agent uses online feedback from the environment and prior knowledge in order to adaptively select an effective policy. Model free approaches address this task by directly mapping external and internal states to actions, while model based methods attempt to construct a model of the environment, followed by a selection of optimal actions based on that model. Given the complementary advantages of both approaches, we suggest a novel algorithm which combines them into a single algorithm, which switches between a model based and a model free mode, depending on the current environmental state and on the status of the agent's knowledge. We prove that such an approach leads to improved perfromance whenever environmental knowledge is available, without compromising perfromance when such knowledge is absent. Numerical simulations demonstrate the effectiveness of the approach and suggest its efficacy in boosting policy gradient learning.\n" + "abstract": "Abstract:In reinforcement learning an agent uses online feedback from the environment and prior knowledge in order to adaptively select an effective policy. Model free approaches address this task by directly mapping external and internal states to actions, while model based methods attempt to construct a model of the environment, followed by a selection of optimal actions based on that model. Given the complementary advantages of both approaches, we suggest a novel algorithm which combines them into a single algorithm, which switches between a model based and a model free mode, depending on the current environmental state and on the status of the agent's knowledge. We prove that such an approach leads to improved performance whenever environmental knowledge is available, without compromising performance when such knowledge is absent. Numerical simulations demonstrate the effectiveness of the approach and suggest its efficacy in boosting policy gradient learning." }, { "author": [ @@ -616,14 +616,14 @@ "Sra, Suvrit" ], "paper_title": "Fast Newton-type Methods for Total Variation Regularization ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Document Analysis Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 313, + "page_form": 313, "page_to": 320, "totle_page": 8, "language": "en", - "abstract": "Abstract:Numerous applications in statistics, signal processing, and machine learning regularize using Total Variation (TV) penalties. We study anisotropic (l1-based) TV and also a related l2-norm variant. We consider for both variants associated (1D) proximity operators, which lead to challenging optimization problems. We solve these problems by developing Newton-type methods that outperfrom the state-of-the-art algorithms. More importantly, our 1D-TV algorithms serve as building blocks for solving the harder task of computing 2- (and higher)-dimensional TV proximity. We illustrate the computational benefits of our methods by applying them to several applications: (i) image denoising; (ii) image deconvolution (by plugging in our TV solvers into publicly available software); and (iii) four variants of fused-lasso. The results show large speedups--and to support our claims, we provide software accompanying this paper.\n" + "abstract": "Abstract:Numerous applications in statistics, signal processing, and machine learning regularize using Total Variation (TV) penalties. We study anisotropic (l1-based) TV and also a related l2-norm variant. We consider for both variants associated (1D) proximity operators, which lead to challenging optimization problems. We solve these problems by developing Newton-type methods that outperform the state-of-the-art algorithms. More importantly, our 1D-TV algorithms serve as building blocks for solving the harder task of computing 2- (and higher)-dimensional TV proximity. We illustrate the computational benefits of our methods by applying them to several applications: (i) image denoising; (ii) image deconvolution (by plugging in our TV solvers into publicly available software); and (iii) four variants of fused-lasso. The results show large speedups--and to support our claims, we provide software accompanying this paper." }, { "author": [ @@ -633,14 +633,14 @@ "Guestrin, Carlos" ], "paper_title": "Parallel Coordinate Descent for L1-Regularized Loss Minimization", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "The European Symposium on Artificial Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 321, + "page_form": 321, "page_to": 328, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose Shotgun, a parallel coordinate descent algorithm for minimizing L1-regularized losses. Though coordinate descent seems inherently sequential, we prove convergence bounds for Shotgun which predict linear speedups, up to a problem-dependent limit. We present a comprehensive empirical study of Shotgun for Lasso and sparse logistic regression. Our theoretical predictions on the potential for parallelism closely match behavior on real data. Shotgun outperfroms other published solvers on a range of large problems, proving to be one of the most scalable algorithms for L1.\n" + "abstract": "Abstract:We propose Shotgun, a parallel coordinate descent algorithm for minimizing L1-regularized losses. Though coordinate descent seems inherently sequential, we prove convergence bounds for Shotgun which predict linear speedups, up to a problem-dependent limit. We present a comprehensive empirical study of Shotgun for Lasso and sparse logistic regression. Our theoretical predictions on the potential for parallelism closely match behavior on real data. Shotgun outperforms other published solvers on a range of large problems, proving to be one of the most scalable algorithms for L1." }, { "author": [ @@ -649,14 +649,14 @@ "Shamir, Ohad" ], "paper_title": "Large-Scale Convex Minimization with a Low-Rank Constraint", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Multiple Classifier Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 329, + "page_form": 329, "page_to": 336, "totle_page": 8, "language": "en", - "abstract": "Abstract:We address the problem of minimizing a convex function over the space of large matrices with low rank. While this optimization problem is hard in general, we propose an efficient greedy algorithm and derive its fromal approximation guarantees. Each iteration of the algorithm involves (approximately) finding the left and right singular vectors corresponding to the largest singular value of a certain matrix, which can be calculated in linear time. This leads to an algorithm which can scale to large matrices arising in several applications such as matrix completion for collaborative filtering and robust low rank matrix approximation.\n" + "abstract": "Abstract:We address the problem of minimizing a convex function over the space of large matrices with low rank. While this optimization problem is hard in general, we propose an efficient greedy algorithm and derive its formal approximation guarantees. Each iteration of the algorithm involves (approximately) finding the left and right singular vectors corresponding to the largest singular value of a certain matrix, which can be calculated in linear time. This leads to an algorithm which can scale to large matrices arising in several applications such as matrix completion for collaborative filtering and robust low rank matrix approximation." }, { "author": [ @@ -664,14 +664,14 @@ "Dunson, David" ], "paper_title": "Approximate Dynamic Programming for Storage Problems ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Joint Conference on Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 337, + "page_form": 337, "page_to": 344, "totle_page": 8, "language": "en", - "abstract": "Abstract:Storage problems are an important subclass of stochastic control problems. This paper presents a new method, approximate dynamic programming for storage, to solve storage problems with continuous, convex decision sets. Unlike other solution procedures, ADPS allows math programming to be used to make decisions each time period, even in the presence of large state variables. We test ADPS on the day ahead wind commitment problem with storage.\n" + "abstract": "Abstract:Storage problems are an important subclass of stochastic control problems. This paper presents a new method, approximate dynamic programming for storage, to solve storage problems with continuous, convex decision sets. Unlike other solution procedures, ADPS allows math programming to be used to make decisions each time period, even in the presence of large state variables. We test ADPS on the day ahead wind commitment problem with storage." }, { "author": [ @@ -679,14 +679,14 @@ "Bilmes, Jeff" ], "paper_title": "Online Submodular Minimization for Combinatorial Structures ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Neural Information Processing", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 345, + "page_form": 345, "page_to": 352, "totle_page": 8, "language": "en", - "abstract": "Abstract:Most results for online decision problems with structured concepts, such as trees or cuts, assume linear costs. In many settings, however, nonlinear costs are more realistic. Owing to their non-separability, these lead to much harder optimization problems. Going beyond linearity, we address online approximation algorithms for structured concepts that allow the cost to be submodular, i.e., nonseparable. In particular, we show regret bounds for three Hannan-consistent strategies that capture different settings. Our results also tighten a regret bound for unconstrained online submodular minimization.\n" + "abstract": "Abstract:Most results for online decision problems with structured concepts, such as trees or cuts, assume linear costs. In many settings, however, nonlinear costs are more realistic. Owing to their non-separability, these lead to much harder optimization problems. Going beyond linearity, we address online approximation algorithms for structured concepts that allow the cost to be submodular, i.e., nonseparable. In particular, we show regret bounds for three Hannan-consistent strategies that capture different settings. Our results also tighten a regret bound for unconstrained online submodular minimization." }, { "author": [ @@ -694,14 +694,14 @@ "Fleet, David" ], "paper_title": "Minimal Loss Hashing for Compact Binary Codes", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "EuroPACS", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 353, + "page_form": 353, "page_to": 360, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose a method for learning similarity-preserving hash functions that map high-dimensional data onto binary codes. The fromulation is based on structured prediction with latent variables and a hinge-like loss function. It is efficient to train for large datasets, scales well to large code lengths, and outperfroms state-of-the-art methods.\n" + "abstract": "Abstract:We propose a method for learning similarity-preserving hash functions that map high-dimensional data onto binary codes. The formulation is based on structured prediction with latent variables and a hinge-like loss function. It is efficient to train for large datasets, scales well to large code lengths, and outperforms state-of-the-art methods." }, { "author": [ @@ -712,14 +712,14 @@ "Carin, Lawrence" ], "paper_title": "The Hierarchical Beta Process for Convolutional Factor Analysis and Deep Learning", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Workshop on Structural and Syntactic Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 361, + "page_form": 361, "page_to": 368, "totle_page": 8, "language": "en", - "abstract": "Abstract:A convolutional factor-analysis model is developed, with the number of filters (factors) inferred via the beta process (BP) and hierarchical BP, for single-task and multi-task learning, respectively. The computation of the model parameters is implemented within a Bayesian setting, employing Gibbs sampling; we explicitly exploit the convolutional nature of the expansion to accelerate computations. The model is used in a multi-level (“deep”) analysis of general data, with specific results presented for image-processing data sets, e.g., classification.\n" + "abstract": "Abstract:A convolutional factor-analysis model is developed, with the number of filters (factors) inferred via the beta process (BP) and hierarchical BP, for single-task and multi-task learning, respectively. The computation of the model parameters is implemented within a Bayesian setting, employing Gibbs sampling; we explicitly exploit the convolutional nature of the expansion to accelerate computations. The model is used in a multi-level (“deep”) analysis of general data, with specific results presented for image-processing data sets, e.g., classification." }, { "author": [ @@ -727,14 +727,14 @@ "Bilmes, Jeff" ], "paper_title": "Simultaneous Learning and Covering with Adversarial Noise ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Colloquium on Grammatical Inference", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 369, + "page_form": 369, "page_to": 376, "totle_page": 8, "language": "en", - "abstract": "Abstract:We study simultaneous learning and covering problems: submodular set cover problems that depend on the solution to an active (query) learning problem. The goal is to jointly minimize the cost of both learning and covering. We extend recent work in this setting to allow for a limited amount of adversarial noise. Certain noisy query learning problems are a special case of our problem. Crucial to our analysis is a lemma showing the logical OR of two submodular cover constraints can be reduced to a single submodular set cover constraint. Combined with known results, this new lemma allows for arbitrary monotone circuits of submodular cover constraints to be reduced to a single constraint. As an example practical application, we present a movie recommendation website that minimizes the total cost of learning what the user wants to watch and recommending a set of movies.\n" + "abstract": "Abstract:We study simultaneous learning and covering problems: submodular set cover problems that depend on the solution to an active (query) learning problem. The goal is to jointly minimize the cost of both learning and covering. We extend recent work in this setting to allow for a limited amount of adversarial noise. Certain noisy query learning problems are a special case of our problem. Crucial to our analysis is a lemma showing the logical OR of two submodular cover constraints can be reduced to a single submodular set cover constraint. Combined with known results, this new lemma allows for arbitrary monotone circuits of submodular cover constraints to be reduced to a single constraint. As an example practical application, we present a movie recommendation website that minimizes the total cost of learning what the user wants to watch and recommending a set of movies." }, { "author": [ @@ -743,14 +743,14 @@ "Carin, Lawrence" ], "paper_title": "Topic Modeling with Nonparametric Markov Tree", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Work-Conference on Artificial and NaturalNeural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 377, + "page_form": 377, "page_to": 384, "totle_page": 8, "language": "en", - "abstract": "Abstract:A new hierarchical tree-based topic model is developed, based on nonparametric Bayesian techniques. The model has two unique attributes: (i) a child node in the tree may have more than one parent, with the goal of eliminating redundant sub-topics deep in the tree; and (ii) parsimonious sub-topics are manifested, by removing redundant usage of words at multiple scales. The depth and width of the tree are unbounded within the prior, with a retrospective sampler employed to adaptively infer the appropriate tree size based upon the corpus under study. Excellent quantitative results are manifested on five standard data sets, and the inferred tree structure is also found to be highly interpretable.\n" + "abstract": "Abstract:A new hierarchical tree-based topic model is developed, based on nonparametric Bayesian techniques. The model has two unique attributes: (i) a child node in the tree may have more than one parent, with the goal of eliminating redundant sub-topics deep in the tree; and (ii) parsimonious sub-topics are manifested, by removing redundant usage of words at multiple scales. The depth and width of the tree are unbounded within the prior, with a retrospective sampler employed to adaptively infer the appropriate tree size based upon the corpus under study. Excellent quantitative results are manifested on five standard data sets, and the inferred tree structure is also found to be highly interpretable." }, { "author": [ @@ -758,14 +758,14 @@ "Neville, Jennifer" ], "paper_title": "Relational Active Learning for Joint Collective Classification Models", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Biometrics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 385, + "page_form": 385, "page_to": 392, "totle_page": 8, "language": "en", - "abstract": "Abstract:In many network domains, labeled data may be costly to acquire---indicating a need for {\\em relational active learning} methods. Recent work has demonstrated that relational model perfromance can be improved by taking network structure into account when choosing instances to label. However, in collective inference settings, {\\em both} model estimation {\\em and} prediction can be improved by acquiring a node's label---since relational models estimate a joint distribution over labels in the network and collective classification methods propagate infromation from labeled training data during prediction. This conflates improvement in learning with improvement in inference, since labeling nodes can reduce inference error without improving the overall quality of the learned model. Here, we use {\\em across-network} classification to separate the effects on learning and prediction, and focus on reduction of learning error. When label propagation is used for learning, we find that labeling based on prediction {\\em certainty} is more effective than labeling based on {\\em uncertainty}. As such, we propose a novel active learning method that combines a network-based {\\em certainty} metric with semi-supervised learning and relational resampling. We evaluate our approach on synthetic and real-world networks and show faster learning compared to several baselines, including the network based method of Bilgic et al. 2010.\n" + "abstract": "Abstract:In many network domains, labeled data may be costly to acquire---indicating a need for {\\em relational active learning} methods. Recent work has demonstrated that relational model performance can be improved by taking network structure into account when choosing instances to label. However, in collective inference settings, {\\em both} model estimation {\\em and} prediction can be improved by acquiring a node's label---since relational models estimate a joint distribution over labels in the network and collective classification methods propagate information from labeled training data during prediction. This conflates improvement in learning with improvement in inference, since labeling nodes can reduce inference error without improving the overall quality of the learned model. Here, we use {\\em across-network} classification to separate the effects on learning and prediction, and focus on reduction of learning error. When label propagation is used for learning, we find that labeling based on prediction {\\em certainty} is more effective than labeling based on {\\em uncertainty}. As such, we propose a novel active learning method that combines a network-based {\\em certainty} metric with semi-supervised learning and relational resampling. We evaluate our approach on synthetic and real-world networks and show faster learning compared to several baselines, including the network based method of Bilgic et al. 2010." }, { "author": [ @@ -773,14 +773,14 @@ "III, Hal Daume" ], "paper_title": "A Co-training Approach for Multi-view Spectral Clustering ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "European Conference on Computational Learning Theory", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 393, + "page_form": 393, "page_to": 400, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose a spectral clustering algorithm for the multi-view setting where we have access to multiple views of the data, each of which can be independently used for clustering. Our spectral clustering algorithm has a flavor of co-training, which is already a widely used idea in semi-supervised learning. We work on the assumption that the true underlying clustering would assign a point to the same cluster irrespective of the view. Hence, we constrain our approach to only search for the clusterings that agree across the views. Our algorithm does not have any hyperparameters to set, which is a major advantage in unsupervised learning. We empirically compare with a number of baseline methods on synthetic and real-world datasets to show the efficacy of the proposed algorithm.\n" + "abstract": "Abstract:We propose a spectral clustering algorithm for the multi-view setting where we have access to multiple views of the data, each of which can be independently used for clustering. Our spectral clustering algorithm has a flavor of co-training, which is already a widely used idea in semi-supervised learning. We work on the assumption that the true underlying clustering would assign a point to the same cluster irrespective of the view. Hence, we constrain our approach to only search for the clusterings that agree across the views. Our algorithm does not have any hyperparameters to set, which is a major advantage in unsupervised learning. We empirically compare with a number of baseline methods on synthetic and real-world datasets to show the efficacy of the proposed algorithm." }, { "author": [ @@ -788,14 +788,14 @@ "Mannor, Shie" ], "paper_title": "Learning from Multiple Outlooks", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Natural Computation", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 401, + "page_form": 401, "page_to": 408, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose a novel problem fromulation of learning a single task when the data are provided in different feature spaces. Each such space is called an outlook, and is assumed to contain both labeled and unlabeled data. The objective is to take advantage of the data from all the outlooks to better classify each of the outlooks. We devise an algorithm that computes optimal affine mappings from different outlooks to a target outlook by matching moments of the empirical distributions. We further derive a probabilistic interpretation of the resulting algorithm and a sample complexity bound indicating how many samples are needed to adequately find the mapping. We report the results of extensive experiments on activity recognition tasks that show the value of the proposed approach in boosting perfromance.\n" + "abstract": "Abstract:We propose a novel problem formulation of learning a single task when the data are provided in different feature spaces. Each such space is called an outlook, and is assumed to contain both labeled and unlabeled data. The objective is to take advantage of the data from all the outlooks to better classify each of the outlooks. We devise an algorithm that computes optimal affine mappings from different outlooks to a target outlook by matching moments of the empirical distributions. We further derive a probabilistic interpretation of the resulting algorithm and a sample complexity bound indicating how many samples are needed to adequately find the mapping. We report the results of extensive experiments on activity recognition tasks that show the value of the proposed approach in boosting performance." }, { "author": [ @@ -804,14 +804,14 @@ "Zheng, Lu" ], "paper_title": "Adaptive Kernel Approximation for Large-Scale Non-Linear SVM Prediction", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Machine Learning and Cybernetics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 409, + "page_form": 409, "page_to": 416, "totle_page": 8, "language": "en", - "abstract": "Abstract:The applicability of non-linear support vector machines (SVMs) has been limited in large-scale data collections because of their linear prediction complexity to the size of support vectors. We propose an efficient prediction algorithm with perfromance guarantee for non-linear SVMs, termed AdaptSVM. It can selectively collapse the kernel function computation to a reduced set of support vectors, compensated by an additional correction term that can be easily computed on-line. It also allows adaptive fall-back to original kernel computation based on its estimated variance and maximum error tolerance. In addition to theoretical analysis, we empirically evaluate on multiple large-scale datasets to show that the proposed algorithm can speed up the prediction process up to 10000 times with only <0.5 accuracy loss.\n" + "abstract": "Abstract:The applicability of non-linear support vector machines (SVMs) has been limited in large-scale data collections because of their linear prediction complexity to the size of support vectors. We propose an efficient prediction algorithm with performance guarantee for non-linear SVMs, termed AdaptSVM. It can selectively collapse the kernel function computation to a reduced set of support vectors, compensated by an additional correction term that can be easily computed on-line. It also allows adaptive fall-back to original kernel computation based on its estimated variance and maximum error tolerance. In addition to theoretical analysis, we empirically evaluate on multiple large-scale datasets to show that the proposed algorithm can speed up the prediction process up to 10000 times with only <0.5 accuracy loss." }, { "author": [ @@ -820,14 +820,14 @@ "Santos-Rodr\\'{i}iguez, Ra\\'{u}l" ], "paper_title": "Risk-Based Generalizations of f-divergences", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Biometric Authentication", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 417, + "page_form": 417, "page_to": 424, "totle_page": 8, "language": "en", - "abstract": "Abstract:We derive a generalized notion of f-divergences, called (f,l)-divergences. We show that this generalization enjoys many of the nice properties of f-divergences, although it is a richer family. It also provides alternative definitions of standard divergences in terms of surrogate risks. As a first practical application of this theory, we derive a new estimator for the Kulback-Leibler divergence that we use for clustering sets of vectors.\n" + "abstract": "Abstract:We derive a generalized notion of f-divergences, called (f,l)-divergences. We show that this generalization enjoys many of the nice properties of f-divergences, although it is a richer family. It also provides alternative definitions of standard divergences in terms of surrogate risks. As a first practical application of this theory, we derive a new estimator for the Kulback-Leibler divergence that we use for clustering sets of vectors." }, { "author": [ @@ -835,14 +835,14 @@ "Lampert, Christoph" ], "paper_title": "Learning Multi-View Neighborhood Preserving Projections ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Automatic Identification Advanced Technologies", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 425, + "page_form": 425, "page_to": 432, "totle_page": 8, "language": "en", - "abstract": "Abstract:We address the problem of metric learning for multi-view data, namely the construction of embedding projections from data in different representations into a shared feature space, such that the Euclidean distance in this space provides a meaningful within-view as well as between-view similarity. Our motivation stems from the problem of cross-media retrieval tasks, where the availability of a joint Euclidean distance function is a prerequisite to allow fast, in particular hashing-based, nearest neighbor queries. We fromulate an objective function that expresses the intuitive concept that matching samples are mapped closely together in the output space, whereas non-matching samples are pushed apart, no matter in which view they are available. The resulting optimization problem is not convex, but it can be decomposed explicitly into a convex and a concave part, thereby allowing efficient optimization using the convex-concave procedure. Experiments on an image retrieval task show that nearest-neighbor based cross-view retrieval is indeed possible, and the proposed technique improves the retrieval accuracy over baseline techniques.\n" + "abstract": "Abstract:We address the problem of metric learning for multi-view data, namely the construction of embedding projections from data in different representations into a shared feature space, such that the Euclidean distance in this space provides a meaningful within-view as well as between-view similarity. Our motivation stems from the problem of cross-media retrieval tasks, where the availability of a joint Euclidean distance function is a prerequisite to allow fast, in particular hashing-based, nearest neighbor queries. We formulate an objective function that expresses the intuitive concept that matching samples are mapped closely together in the output space, whereas non-matching samples are pushed apart, no matter in which view they are available. The resulting optimization problem is not convex, but it can be decomposed explicitly into a convex and a concave part, thereby allowing efficient optimization using the convex-concave procedure. Experiments on an image retrieval task show that nearest-neighbor based cross-view retrieval is indeed possible, and the proposed technique improves the retrieval accuracy over baseline techniques." }, { "author": [ @@ -850,14 +850,14 @@ "Cesa-Bianchi, Nicol\\`{o}" ], "paper_title": "Better Algorithms for Selective Sampling ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Indian Conference on Computer Vision, Graphics & Image Processing", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 433, + "page_form": 433, "page_to": 440, "totle_page": 8, "language": "en", - "abstract": "Abstract:We study online algorithms for selective sampling that use regularized least squares (RLS) as base classifier. These algorithms typically perfrom well in practice, and some of them have fromal guarantees on their mistake and query rates. We refine and extend these guarantees in various ways, proposing algorithmic variants that exhibit better empirical behavior while enjoying perfromance guarantees under much more general conditions. We also show a simple way of coupling a generic gradient-based classifier with a specific RLS-based selective sampler, obtaining hybrid algorithms with combined perfromance guarantees.\n" + "abstract": "Abstract:We study online algorithms for selective sampling that use regularized least squares (RLS) as base classifier. These algorithms typically perform well in practice, and some of them have formal guarantees on their mistake and query rates. We refine and extend these guarantees in various ways, proposing algorithmic variants that exhibit better empirical behavior while enjoying performance guarantees under much more general conditions. We also show a simple way of coupling a generic gradient-based classifier with a specific RLS-based selective sampler, obtaining hybrid algorithms with combined performance guarantees." }, { "author": [ @@ -865,14 +865,14 @@ "Cl\\'{e}mencon, St\\'{e}phan" ], "paper_title": "Minimax Learning Rates for Bipartite Ranking and Plug-in Rules", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Support Vector Machines", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 441, + "page_form": 441, "page_to": 448, "totle_page": 8, "language": "en", - "abstract": "Abstract:While it is now well-known in the standard binary classi\fcation setup, that, under suitable margin assumptions and complexity conditions on the regression function, fast or even super-fast rates (i.e. rates faster than n^(-1/2) or even faster than n^-1) can be achieved by plug-in classi\fers, no result of this nature has been proved yet in the context of bipartite ranking, though akin to that of classi\fcation. It is the main purpose of the present paper to investigate this issue. Viewing bipartite ranking as a nested continuous collection of cost-sensitive classi\fcation problems, we exhibit a global low noise condition under which certain plug-in ranking rules can be shown to achieve fast (but not super-fast) rates, establishing thus minimax upper bounds for the excess of ranking risk.\n" + "abstract": "Abstract:While it is now well-known in the standard binary classi\fcation setup, that, under suitable margin assumptions and complexity conditions on the regression function, fast or even super-fast rates (i.e. rates faster than n^(-1/2) or even faster than n^-1) can be achieved by plug-in classi\fers, no result of this nature has been proved yet in the context of bipartite ranking, though akin to that of classi\fcation. It is the main purpose of the present paper to investigate this issue. Viewing bipartite ranking as a nested continuous collection of cost-sensitive classi\fcation problems, we exhibit a global low noise condition under which certain plug-in ranking rules can be shown to achieve fast (but not super-fast) rates, establishing thus minimax upper bounds for the excess of ranking risk." }, { "author": [ @@ -880,14 +880,14 @@ "Toussaint, Marc" ], "paper_title": "Task Space Retrieval Using Inverse Feedback Control ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "ICCI(ieee) - International Conference on Cognitive Informatics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 449, + "page_form": 449, "page_to": 456, "totle_page": 8, "language": "en", - "abstract": "Abstract:Learning complex skills by repeating and generalizing expert behavior is a fundamental problem in robotics. A common approach is learning from demonstration: given examples of correct motions, learn a policy mapping state to action consistent with the training data. However, the usual approaches do not answer the question of what are appropriate representations to generate motions for specific tasks. Inspired by Inverse Optimal Control, we present a novel method to learn latent costs, imitate and generalize demonstrated behavior, and discover a task relevant motion representation: Task Space Retrieval Using Inverse Feedback Control (TRIC). We use the learned latent costs to create motion with a feedback controller. We tested our method on robot grasping of objects, a challenging high-dimensional task. TRIC learns the important control dimensions for the grasping task from a few example movements and is able to robustly approach and grasp objects in new situations.\n" + "abstract": "Abstract:Learning complex skills by repeating and generalizing expert behavior is a fundamental problem in robotics. A common approach is learning from demonstration: given examples of correct motions, learn a policy mapping state to action consistent with the training data. However, the usual approaches do not answer the question of what are appropriate representations to generate motions for specific tasks. Inspired by Inverse Optimal Control, we present a novel method to learn latent costs, imitate and generalize demonstrated behavior, and discover a task relevant motion representation: Task Space Retrieval Using Inverse Feedback Control (TRIC). We use the learned latent costs to create motion with a feedback controller. We tested our method on robot grasping of objects, a challenging high-dimensional task. TRIC learns the important control dimensions for the grasping task from a few example movements and is able to robustly approach and grasp objects in new situations." }, { "author": [ @@ -896,14 +896,14 @@ "Kaski, Samuel" ], "paper_title": "Bayesian CCA via Group Sparsity", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Machine Learning and Data Mining in Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 457, + "page_form": 457, "page_to": 464, "totle_page": 8, "language": "en", - "abstract": "Abstract:Bayesian treatments of Canonical Correlation Analysis (CCA) -type latent variable models have been recently proposed for coping with overfitting in small sample sizes, as well as for producing factorizations of the data sources into correlated and non-shared effects. However, all of the current implementations of Bayesian CCA and its extensions are computationally inefficient for high-dimensional data and, as shown in this paper, break down completely for high-dimensional sources with low sample count. Furthermore, they cannot reliably separate the correlated effects from non-shared ones. We propose a new Bayesian CCA variant that is computationally efficient and works for high-dimensional data, while also learning the factorization more accurately. The improvements are gained by introducing a group sparsity assumption and an improved variational approximation. The method is demonstrated to work well on multi-label prediction tasks and in analyzing brain correlates of naturalistic audio stimulation.\n" + "abstract": "Abstract:Bayesian treatments of Canonical Correlation Analysis (CCA) -type latent variable models have been recently proposed for coping with overfitting in small sample sizes, as well as for producing factorizations of the data sources into correlated and non-shared effects. However, all of the current implementations of Bayesian CCA and its extensions are computationally inefficient for high-dimensional data and, as shown in this paper, break down completely for high-dimensional sources with low sample count. Furthermore, they cannot reliably separate the correlated effects from non-shared ones. We propose a new Bayesian CCA variant that is computationally efficient and works for high-dimensional data, while also learning the factorization more accurately. The improvements are gained by introducing a group sparsity assumption and an improved variational approximation. The method is demonstrated to work well on multi-label prediction tasks and in analyzing brain correlates of naturalistic audio stimulation." }, { "author": [ @@ -911,14 +911,14 @@ "Rasmussen, Carl" ], "paper_title": "PILCO: A Model-Based and Data-Efficient Approach to Policy Search ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Advances in Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 465, + "page_form": 465, "page_to": 472, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper, we introduce PILCO, a practical, data-efficient model-based policy search method. PILCO reduces model bias, one of the key problems of model-based reinforcement learning, in a principled way. By learning a probabilistic dynamics model and explicitly incorporating model uncertainty into long-term planning, PILCO can cope with very little data and facilitates learning from scratch in only a few trials. Policy evaluation is perfromed in closed from using state-of-the-art approximate inference. Furthermore, policy gradients are computed analytically for policy improvement. We report unprecedented learning efficiency on challenging and high-dimensional control tasks.\n" + "abstract": "Abstract:In this paper, we introduce PILCO, a practical, data-efficient model-based policy search method. PILCO reduces model bias, one of the key problems of model-based reinforcement learning, in a principled way. By learning a probabilistic dynamics model and explicitly incorporating model uncertainty into long-term planning, PILCO can cope with very little data and facilitates learning from scratch in only a few trials. Policy evaluation is performed in closed form using state-of-the-art approximate inference. Furthermore, policy gradients are computed analytically for policy improvement. We report unprecedented learning efficiency on challenging and high-dimensional control tasks." }, { "author": [ @@ -926,14 +926,14 @@ "Takeuchi, Ichiro" ], "paper_title": "Suboptimal Solution Path Algorithm for Support Vector Machine ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Graph Based Representations in Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 473, + "page_form": 473, "page_to": 480, "totle_page": 8, "language": "en", - "abstract": "Abstract:We consider a suboptimal solution path algorithm for the Support Vector Machine. The solution path algorithm is known as an effective tool for solving a sequence of a parametrized optimization problems in machine learning. However, the algorithm needs to keep strict optimality conditions satisfied everywhere on the path. This requirement narrows the applicability of the path algorithm and adversely affects its computational efficiency. In our algorithm, user can specify tolerances to the optimality and control the trade-off between accuracy of the solution and the computational cost. We also show that our suboptimal solutions can be interpreted as the solution of a perturbed optimization problem from the original one, provide some theoretical analyses of our algorithm based on a novel interpretation. The experimental results demonstrate the effectiveness of our algorithm in terms of efficiency and accuracy.\n" + "abstract": "Abstract:We consider a suboptimal solution path algorithm for the Support Vector Machine. The solution path algorithm is known as an effective tool for solving a sequence of a parametrized optimization problems in machine learning. However, the algorithm needs to keep strict optimality conditions satisfied everywhere on the path. This requirement narrows the applicability of the path algorithm and adversely affects its computational efficiency. In our algorithm, user can specify tolerances to the optimality and control the trade-off between accuracy of the solution and the computational cost. We also show that our suboptimal solutions can be interpreted as the solution of a perturbed optimization problem from the original one, provide some theoretical analyses of our algorithm based on a novel interpretation. The experimental results demonstrate the effectiveness of our algorithm in terms of efficiency and accuracy." }, { "author": [ @@ -943,14 +943,14 @@ "Schmidhuber, J\\\"{u}rgen" ], "paper_title": "Incremental Basis Construction from Temporal Difference Error ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Stochastic Algorithms", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 481, + "page_form": 481, "page_to": 488, "totle_page": 8, "language": "en", - "abstract": "Abstract:In many reinforcement-learning (RL) systems, the value function is approximated as a linear combination of a fixed set of basis functions. Perfromance can be improved by adding to this set. Previous approaches construct a series of basis functions that in sufficient number can eventually represent the value function. In contrast, we show that there is a single, ideal basis function, which can directly represent the value function. Its addition to the set immediately reduces the error to zero---without changing existing weights. Moreover, this ideal basis function is simply the value function that results from replacing the MDP's reward function with its Bellman error. This result suggests a novel method for improving value-function estimation: a primary reinforcement learner estimates its value function using its present basis functions; it then sends its TD error to a secondary learner, which interprets that error as a reward function and estimates the corresponding value function; the resulting value function then becomes the primary learner's new basis function. We present both batch and online versions in combination with incremental basis projection, and demonstrate that the perfromance is superior to existing methods, especially in the case of large discount factors.\n" + "abstract": "Abstract:In many reinforcement-learning (RL) systems, the value function is approximated as a linear combination of a fixed set of basis functions. Performance can be improved by adding to this set. Previous approaches construct a series of basis functions that in sufficient number can eventually represent the value function. In contrast, we show that there is a single, ideal basis function, which can directly represent the value function. Its addition to the set immediately reduces the error to zero---without changing existing weights. Moreover, this ideal basis function is simply the value function that results from replacing the MDP's reward function with its Bellman error. This result suggests a novel method for improving value-function estimation: a primary reinforcement learner estimates its value function using its present basis functions; it then sends its TD error to a secondary learner, which interprets that error as a reward function and estimates the corresponding value function; the resulting value function then becomes the primary learner's new basis function. We present both batch and online versions in combination with incremental basis projection, and demonstrate that the performance is superior to existing methods, especially in the case of large discount factors." }, { "author": [ @@ -958,14 +958,14 @@ "Blei, David" ], "paper_title": "Predicting Legislative Roll Calls from Text ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Intelligent Computing", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 489, + "page_form": 489, "page_to": 496, "totle_page": 8, "language": "en", - "abstract": "Abstract:We develop several predictive models linking legislative sentiment to legislative text. Our models, which draw on ideas from ideal point estimation and topic models, predict voting patterns based on the contents of bills and infer the political leanings of legislators. With supervised topics, we provide an exploratory window into how the language of the law is correlated with political support. We also derive approximate posterior inference algorithms based on variational methods. Across 12 years of legislative data, we predict specific voting patterns with high accuracy.\n" + "abstract": "Abstract:We develop several predictive models linking legislative sentiment to legislative text. Our models, which draw on ideas from ideal point estimation and topic models, predict voting patterns based on the contents of bills and infer the political leanings of legislators. With supervised topics, we provide an exploratory window into how the language of the law is correlated with political support. We also derive approximate posterior inference algorithms based on variational methods. Across 12 years of legislative data, we predict specific voting patterns with high accuracy." }, { "author": [ @@ -974,28 +974,28 @@ "Babacan, Derin" ], "paper_title": "On Bayesian PCA: Automatic Dimensionality Selection and Analytic Solution", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Bildverarbeitung fur die Medizin", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 497, + "page_form": 497, "page_to": 504, "totle_page": 8, "language": "en", - "abstract": "Abstract:In probabilistic PCA, the fully Bayesian estimation is computationally intractable. To cope with this problem, two types of approximation schemes were introduced: the partially Bayesian PCA (PB-PCA) where only the latent variables are integrated out, and the variational Bayesian PCA (VB-PCA) where the loading vectors are also integrated out. The VB-PCA was proposed as an improved variant of PB-PCA for enabling automatic dimensionality selection (ADS). In this paper, we investigate whether VB-PCA is really the best choice from the viewpoints of computational efficiency and ADS. We first show that ADS is not the unique feature of VB-PCA---PB-PCA is also actually equipped with ADS. We further show that PB-PCA is more advantageous in computational efficiency than VB-PCA because the global solution of PB-PCA can be computed analytically. However, we also show the negative fact that PB-PCA results in a trivial solution in the empirical Bayesian framework. We next consider a simplified variant of VB-PCA, where the latent variables and loading vectors are assumed to be mutually independent (while the ordinary VB-PCA only requires matrix-wise independence). We show that this simplified VB-PCA is the most advantageous in practice because its empirical Bayes solution experimentally works as well as the original VB-PCA, and its global optimal solution can be computed efficiently in a closed from.\n" + "abstract": "Abstract:In probabilistic PCA, the fully Bayesian estimation is computationally intractable. To cope with this problem, two types of approximation schemes were introduced: the partially Bayesian PCA (PB-PCA) where only the latent variables are integrated out, and the variational Bayesian PCA (VB-PCA) where the loading vectors are also integrated out. The VB-PCA was proposed as an improved variant of PB-PCA for enabling automatic dimensionality selection (ADS). In this paper, we investigate whether VB-PCA is really the best choice from the viewpoints of computational efficiency and ADS. We first show that ADS is not the unique feature of VB-PCA---PB-PCA is also actually equipped with ADS. We further show that PB-PCA is more advantageous in computational efficiency than VB-PCA because the global solution of PB-PCA can be computed analytically. However, we also show the negative fact that PB-PCA results in a trivial solution in the empirical Bayesian framework. We next consider a simplified variant of VB-PCA, where the latent variables and loading vectors are assumed to be mutually independent (while the ordinary VB-PCA only requires matrix-wise independence). We show that this simplified VB-PCA is the most advantageous in practice because its empirical Bayes solution experimentally works as well as the original VB-PCA, and its global optimal solution can be computed efficiently in a closed form." }, { "author": [ "Bylander, Tom" ], "paper_title": "Learning Linear Functions with Quadratic and Linear Multiplicative Updates", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Artificial Intelligence and Soft Computing", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 505, + "page_form": 505, "page_to": 512, "totle_page": 8, "language": "en", - "abstract": "Abstract:We analyze variations of multiplicative updates for learning linear functions online. These can be described as substituting exponentiation in the Exponentiated Gradient (EG) algorithm with quadratic and linear functions. Both kinds of updates substitute exponentiation with simpler operations and reduce dependence on the parameter that specifies the sum of the weights during learning. In particular, the linear multiplicative update places no restrictions on the sum of the weights, and, under a wide range of conditions, achieves worst-case behavior close to the EG algorithm. We perfrom our analysis for square loss and absolute loss, and for regression and classification. We also describe some experiments showing that the perfromance of our algorithms are comparable to EG and the $p$-norm algorithm.\n" + "abstract": "Abstract:We analyze variations of multiplicative updates for learning linear functions online. These can be described as substituting exponentiation in the Exponentiated Gradient (EG) algorithm with quadratic and linear functions. Both kinds of updates substitute exponentiation with simpler operations and reduce dependence on the parameter that specifies the sum of the weights during learning. In particular, the linear multiplicative update places no restrictions on the sum of the weights, and, under a wide range of conditions, achieves worst-case behavior close to the EG algorithm. We perform our analysis for square loss and absolute loss, and for regression and classification. We also describe some experiments showing that the performance of our algorithms are comparable to EG and the $p$-norm algorithm." }, { "author": [ @@ -1004,14 +1004,14 @@ "Bengio, Yoshua" ], "paper_title": "Domain Adaptation for Large-Scale Sentiment Classification: A Deep Learning Approach", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Brazilian Symposium on Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 513, + "page_form": 513, "page_to": 520, "totle_page": 8, "language": "en", - "abstract": "Abstract:The exponential increase in the availability of online reviews and recommendations makes sentiment classification an interesting topic in academic and industrial research. Reviews can span so many different domains that it is difficult to gather annotated training data for all of them. Hence, this paper studies the problem of domain adaptation for sentiment classifiers, hereby a system is trained on labeled reviews from one source domain but is meant to be deployed on another. We propose a deep learning approach which learns to extract a meaningful representation for each review in an unsupervised fashion. Sentiment classifiers trained with this high-level feature representation clearly outperfrom state-of-the-art methods on a benchmark composed of reviews of 4 types of Amazon products. Furthermore, this method scales well and allowed us to successfully perfrom domain adaptation on a larger industrial-strength dataset of 22 domains.\n" + "abstract": "Abstract:The exponential increase in the availability of online reviews and recommendations makes sentiment classification an interesting topic in academic and industrial research. Reviews can span so many different domains that it is difficult to gather annotated training data for all of them. Hence, this paper studies the problem of domain adaptation for sentiment classifiers, hereby a system is trained on labeled reviews from one source domain but is meant to be deployed on another. We propose a deep learning approach which learns to extract a meaningful representation for each review in an unsupervised fashion. Sentiment classifiers trained with this high-level feature representation clearly outperform state-of-the-art methods on a benchmark composed of reviews of 4 types of Amazon products. Furthermore, this method scales well and allowed us to successfully perform domain adaptation on a larger industrial-strength dataset of 22 domains." }, { "author": [ @@ -1020,28 +1020,28 @@ "Sha, Fei" ], "paper_title": "Learning with Whom to Share in Multi-task Feature Learning ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Enformatika Conferences", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 521, + "page_form": 521, "page_to": 528, "totle_page": 8, "language": "en", - "abstract": "Abstract:In multi-task learning (MTL), multiple tasks are learnt jointly. A major assumption for this paradigm is that all those tasks are indeed related so that the joint training is appropriate and beneficial. In this paper, we study the problem of multi-task learning of shared feature representations among tasks, while simultaneously determining ``with whom'' each task should share. We fromulate the problem as a mixed integer programming and provide an alternating minimization technique to solve the optimization problem of jointly identifying grouping structures and parameters. The algorithm monotonically decreases the objective function and converges to a local optimum. Compared to the standard MTL paradigm where all tasks are in a single group, our algorithm improves its perfromance with statistical significance for three out of the four datasets we have studied. We also demonstrate its advantage over other task grouping techniques investigated in literature.\n" + "abstract": "Abstract:In multi-task learning (MTL), multiple tasks are learnt jointly. A major assumption for this paradigm is that all those tasks are indeed related so that the joint training is appropriate and beneficial. In this paper, we study the problem of multi-task learning of shared feature representations among tasks, while simultaneously determining ``with whom'' each task should share. We formulate the problem as a mixed integer programming and provide an alternating minimization technique to solve the optimization problem of jointly identifying grouping structures and parameters. The algorithm monotonically decreases the objective function and converges to a local optimum. Compared to the standard MTL paradigm where all tasks are in a single group, our algorithm improves its performance with statistical significance for three out of the four datasets we have studied. We also demonstrate its advantage over other task grouping techniques investigated in literature." }, { "author": [ "Reyzin, Lev" ], "paper_title": "Boosting on a Budget: Sampling for Feature-Efficient Prediction ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Reading and Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 529, + "page_form": 529, "page_to": 536, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper, we tackle the problem of feature-efficient prediction: classification using a limited number of features per test example. We show that modifying an ensemble classifier such as AdaBoost, by sampling hypotheses from its final weighted predictor, is well-suited for this task. We further consider an extension of this problem, where the costs of examining the various features can differ from one another, and we give an algorithm for this more general setting. We prove the correctness of our algorithms and derive bounds for the number of samples needed for given error rates. We also experimentally verify the effectiveness of our methods.\n" + "abstract": "Abstract:In this paper, we tackle the problem of feature-efficient prediction: classification using a limited number of features per test example. We show that modifying an ensemble classifier such as AdaBoost, by sampling hypotheses from its final weighted predictor, is well-suited for this task. We further consider an extension of this problem, where the costs of examining the various features can differ from one another, and we give an algorithm for this more general setting. We prove the correctness of our algorithms and derive bounds for the number of samples needed for given error rates. We also experimentally verify the effectiveness of our methods." }, { "author": [ @@ -1051,14 +1051,14 @@ "Dzeroski, Saso" ], "paper_title": "Speeding-Up Hoeffding-Based Regression Trees With Options ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Iberoamerican Congress on Pattern Recognition CIARP", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 537, + "page_form": 537, "page_to": 544, "totle_page": 8, "language": "en", - "abstract": "Abstract:Data streams are ubiquitous and have in the last two decades become an important research topic. For their predictive non-parametric analysis, Hoeffding-based trees are often a method of choice, offering a possibility of any-time predictions. However, one of their main problems is the delay in learning progress due to the existence of equally discriminative attributes. Options are a natural way to deal with this problem. Option trees build upon regular trees by adding splitting options in the internal nodes. As such they are known to improve accuracy, stability and reduce ambiguity. In this paper, we present on-line option trees for faster learning on numerical data streams. Our results show that options improve the any-time perfromance of ordinary on-line regression trees, while preserving the interpretable structure of trees and without significantly increasing the computational complexity of the algorithm.\n" + "abstract": "Abstract:Data streams are ubiquitous and have in the last two decades become an important research topic. For their predictive non-parametric analysis, Hoeffding-based trees are often a method of choice, offering a possibility of any-time predictions. However, one of their main problems is the delay in learning progress due to the existence of equally discriminative attributes. Options are a natural way to deal with this problem. Option trees build upon regular trees by adding splitting options in the internal nodes. As such they are known to improve accuracy, stability and reduce ambiguity. In this paper, we present on-line option trees for faster learning on numerical data streams. Our results show that options improve the any-time performance of ordinary on-line regression trees, while preserving the interpretable structure of trees and without significantly increasing the computational complexity of the algorithm." }, { "author": [ @@ -1067,14 +1067,14 @@ "Sepulchre, Rodolphe" ], "paper_title": "Linear Regression under Fixed-Rank Constraints: A Riemannian Approach", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Pattern Recognition and Machine Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 545, + "page_form": 545, "page_to": 552, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper, we tackle the problem of learning a linear regression model whose parameter is a fixed-rank matrix. We study the Riemannian manifold geometry of the set of fixed-rank matrices and develop efficient line-search algorithms. The proposed algorithms have many applications, scale to high-dimensional problems, enjoy local convergence properties and confer a geometric basis to recent contributions on learning fixed-rank matrices. Numerical experiments on benchmarks suggest that the proposed algorithms compete with the state-of-the-art, and that manifold optimization offers a versatile framework for the design of rank-constrained machine learning algorithms.\n" + "abstract": "Abstract:In this paper, we tackle the problem of learning a linear regression model whose parameter is a fixed-rank matrix. We study the Riemannian manifold geometry of the set of fixed-rank matrices and develop efficient line-search algorithms. The proposed algorithms have many applications, scale to high-dimensional problems, enjoy local convergence properties and confer a geometric basis to recent contributions on learning fixed-rank matrices. Numerical experiments on benchmarks suggest that the proposed algorithms compete with the state-of-the-art, and that manifold optimization offers a versatile framework for the design of rank-constrained machine learning algorithms." }, { "author": [ @@ -1084,14 +1084,14 @@ "Huang, Heng" ], "paper_title": "Cauchy Graph Embedding", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Pattern Recognition in Information Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 553, + "page_form": 553, "page_to": 560, "totle_page": 8, "language": "en", - "abstract": "Abstract:Laplacian embedding provides a low-dimensional representation for the nodes of a graph where the edge weights denote pairwise similarity among the node objects. It is commonly assumed that the Laplacian embedding results preserve the local topology of the original data on the low-dimensional projected subspaces, i.e., for any pair of graph nodes with large similarity, they should be embedded closely in the embedded space. However, in this paper, we will show that the Laplacian embedding often cannot preserve local topology well as we expected. To enhance the local topology preserving property in graph embedding, we propose a novel Cauchy graph embedding which preserves the similarity relationships of the original data in the embedded space via a new objective. Consequentially the machine learning tasks (such as k Nearest Neighbor type classifications) can be easily conducted on the embedded data with better perfromance. The experimental results on both synthetic and real world benchmark data sets demonstrate the usefulness of this new type of embedding.\n" + "abstract": "Abstract:Laplacian embedding provides a low-dimensional representation for the nodes of a graph where the edge weights denote pairwise similarity among the node objects. It is commonly assumed that the Laplacian embedding results preserve the local topology of the original data on the low-dimensional projected subspaces, i.e., for any pair of graph nodes with large similarity, they should be embedded closely in the embedded space. However, in this paper, we will show that the Laplacian embedding often cannot preserve local topology well as we expected. To enhance the local topology preserving property in graph embedding, we propose a novel Cauchy graph embedding which preserves the similarity relationships of the original data in the embedded space via a new objective. Consequentially the machine learning tasks (such as k Nearest Neighbor type classifications) can be easily conducted on the embedded data with better performance. The experimental results on both synthetic and real world benchmark data sets demonstrate the usefulness of this new type of embedding." }, { "author": [ @@ -1100,14 +1100,14 @@ "Sch\\\"{o}lkopf, Bernhard" ], "paper_title": "Uncovering the Temporal Dynamics of Diffusion Networks ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IEEE Symposium on Computational Intelligence and Data Mining", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 561, + "page_form": 561, "page_to": 568, "totle_page": 8, "language": "en", - "abstract": "Abstract:Time plays an essential role in the diffusion of infromation, influence and disease over networks. In many cases we only observe when a node copies infromation, makes a decision or becomes infected -- but the connectivity, transmission rates between nodes and transmission sources are unknown. Inferring the underlying dynamics is of outstanding interest since it enables forecasting, influencing and retarding infections, broadly construed. To this end, we model diffusion processes as discrete networks of continuous temporal processes occurring at different rates. Given cascade data -- observed infection times of nodes -- we infer the edges of the global diffusion network and estimate the transmission rates of each edge that best explain the observed data. The optimization problem is convex. The model naturally (without heuristics) imposes sparse solutions and requires no parameter tuning. The problem decouples into a collection of independent smaller problems, thus scaling easily to networks on the order of hundreds of thousands of nodes. Experiments on real and synthetic data show that our algorithm both recovers the edges of diffusion networks and accurately estimates their transmission rates from cascade data.\n" + "abstract": "Abstract:Time plays an essential role in the diffusion of information, influence and disease over networks. In many cases we only observe when a node copies information, makes a decision or becomes infected -- but the connectivity, transmission rates between nodes and transmission sources are unknown. Inferring the underlying dynamics is of outstanding interest since it enables forecasting, influencing and retarding infections, broadly construed. To this end, we model diffusion processes as discrete networks of continuous temporal processes occurring at different rates. Given cascade data -- observed infection times of nodes -- we infer the edges of the global diffusion network and estimate the transmission rates of each edge that best explain the observed data. The optimization problem is convex. The model naturally (without heuristics) imposes sparse solutions and requires no parameter tuning. The problem decouples into a collection of independent smaller problems, thus scaling easily to networks on the order of hundreds of thousands of nodes. Experiments on real and synthetic data show that our algorithm both recovers the edges of diffusion networks and accurately estimates their transmission rates from cascade data." }, { "author": [ @@ -1115,14 +1115,14 @@ "Koller, Daphne" ], "paper_title": "Multiclass Boosting with Hinge Loss based on Output Coding ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Summer School on Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 569, + "page_form": 569, "page_to": 576, "totle_page": 8, "language": "en", - "abstract": "Abstract:Multiclass classification is an important and fundamental problem in machine learning. A popular family of multiclass classification methods belongs to reducing multiclass to binary based on output coding. Several multiclass boosting algorithms have been proposed to learn the coding matrix and the associated binary classifiers in a problem-dependent way. These algorithms can be unified under a sum-of-exponential loss function defined in the domain of margins. Instead, multiclass SVM uses another type of loss function based on hinge loss. In this paper, we present a new output-coding-based multiclass boosting algorithm using the multiclass hinge loss, which we call HingeBoost.OC. HingeBoost.OC is tested on various real world datasets and shows better perfromance than the existing multiclass boosting algorithm AdaBoost.ERP, one-vs-one, one-vs-all, ECOC and multiclass SVM in a majority of different cases.\n" + "abstract": "Abstract:Multiclass classification is an important and fundamental problem in machine learning. A popular family of multiclass classification methods belongs to reducing multiclass to binary based on output coding. Several multiclass boosting algorithms have been proposed to learn the coding matrix and the associated binary classifiers in a problem-dependent way. These algorithms can be unified under a sum-of-exponential loss function defined in the domain of margins. Instead, multiclass SVM uses another type of loss function based on hinge loss. In this paper, we present a new output-coding-based multiclass boosting algorithm using the multiclass hinge loss, which we call HingeBoost.OC. HingeBoost.OC is tested on various real world datasets and shows better performance than the existing multiclass boosting algorithm AdaBoost.ERP, one-vs-one, one-vs-all, ECOC and multiclass SVM in a majority of different cases." }, { "author": [ @@ -1130,14 +1130,14 @@ "Bilmes, Jeff" ], "paper_title": "Approximation Bounds for Inference using Cooperative Cuts ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Workshop on ML", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 577, + "page_form": 577, "page_to": 584, "totle_page": 8, "language": "en", - "abstract": "Abstract:We analyze a family of probability distributions that are characterized by an embedded combinatorial structure. This family includes models having arbitrary treewidth and arbitrary sized factors. Unlike general models with such freedom, where the “most probable explanation” (MPE) problem is inapproximable, the combinatorial structure within our model, in particular the indirect use of submodularity, leads to several MPE algorithms that all have approximation guarantees.\n" + "abstract": "Abstract:We analyze a family of probability distributions that are characterized by an embedded combinatorial structure. This family includes models having arbitrary treewidth and arbitrary sized factors. Unlike general models with such freedom, where the “most probable explanation” (MPE) problem is inapproximable, the combinatorial structure within our model, in particular the indirect use of submodularity, leads to several MPE algorithms that all have approximation guarantees." }, { "author": [ @@ -1145,15 +1145,15 @@ "Flach, Peter", "Ferri, Cèsar" ], - "paper_title": "Brier Curves: a New Cost-Based Visualisation of Classifier Perfromance", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "paper_title": "Brier Curves: a New Cost-Based Visualisation of Classifier Performance", + "booktitle": "International Workshop on Combinatorial Image Analysis", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 585, + "page_form": 585, "page_to": 592, "totle_page": 8, "language": "en", - "abstract": "Abstract:It is often necessary to evaluate classifier perfromance over a range of operating conditions, rather than as a point estimate. This is typically assessed through the construction of ‘curves’over a ‘space’, visualising how one or two perfromance metrics vary with the operating condition. For binary classifiers in particular, cost space is a natural way of showing this range of perfromance, visualising loss against operating condition. However, the curves which have been traditionally drawn in cost space, known as cost curves, show the optimal loss, and hence assume knowledge of the optimal decision threshold for a given operating condition. Clearly, this leads to an optimistic assessment of classifier perfromance. In this paper we propose a more natural way of visualising classifier perfromance in cost space, which is to plot probabilistic loss on the y-axis, i.e., the loss arising from the probability estimates. This new curve provides new ways of understanding classifier perfromance and new tools to compare classifiers. In addition, we show that the area under this curve is exactly the Brier score, one of the most popular perfromance metrics for probabilistic classifiers.\n" + "abstract": "Abstract:It is often necessary to evaluate classifier performance over a range of operating conditions, rather than as a point estimate. This is typically assessed through the construction of ‘curves’over a ‘space’, visualising how one or two performance metrics vary with the operating condition. For binary classifiers in particular, cost space is a natural way of showing this range of performance, visualising loss against operating condition. However, the curves which have been traditionally drawn in cost space, known as cost curves, show the optimal loss, and hence assume knowledge of the optimal decision threshold for a given operating condition. Clearly, this leads to an optimistic assessment of classifier performance. In this paper we propose a more natural way of visualising classifier performance in cost space, which is to plot probabilistic loss on the y-axis, i.e., the loss arising from the probability estimates. This new curve provides new ways of understanding classifier performance and new tools to compare classifiers. In addition, we show that the area under this curve is exactly the Brier score, one of the most popular performance metrics for probabilistic classifiers." }, { "author": [ @@ -1162,14 +1162,14 @@ "Szafranski, Marie" ], "paper_title": "Semi-supervised Penalized Output Kernel Regression for Link Prediction ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "ICSC Symposium on Neural Computation", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 593, + "page_form": 593, "page_to": 600, "totle_page": 8, "language": "en", - "abstract": "Abstract:Link prediction is addressed as an output kernel learning task through semi-supervised Output Kernel Regression. Working in the framework of RKHS theory with vector-valued functions, we establish a new representer theorem devoted to semi-supervised least square regression. We then apply it to get a new model (POKR: Penalized Output Kernel Regression) and show its relevance using numerical experiments on artificial networks and two real applications using a very low percentage of labeled data in a transductive setting.\n" + "abstract": "Abstract:Link prediction is addressed as an output kernel learning task through semi-supervised Output Kernel Regression. Working in the framework of RKHS theory with vector-valued functions, we establish a new representer theorem devoted to semi-supervised least square regression. We then apply it to get a new model (POKR: Penalized Output Kernel Regression) and show its relevance using numerical experiments on artificial networks and two real applications using a very low percentage of labeled data in a transductive setting." }, { "author": [ @@ -1177,14 +1177,14 @@ "Sirotkin, Alexander" ], "paper_title": "A New Bayesian Rating System for Team Competitions ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Artificial Neural Networks in Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 601, + "page_form": 601, "page_to": 608, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present a novel probabilistic rating system for team competitions. Building upon TrueSkill(tm), we change the factor graph structure to cope with the problems of TrueSkill(tm), e.g., multiway ties and variable team size. We give detailed inference algorithms for the new structure. Experimental results show a significant improvement over TrueSkill(tm).\n" + "abstract": "Abstract:We present a novel probabilistic rating system for team competitions. Building upon TrueSkill(tm), we change the factor graph structure to cope with the problems of TrueSkill(tm), e.g., multiway ties and variable team size. We give detailed inference algorithms for the new structure. Experimental results show a significant improvement over TrueSkill(tm)." }, { "author": [ @@ -1199,14 +1199,14 @@ "Olukotun, Kunle" ], "paper_title": "OptiML: An Implicitly Parallel Domain-Specific Language for Machine Learning ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Emergent Neural Computational Architectures Based on Neuroscience", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 609, + "page_form": 609, "page_to": 616, "totle_page": 8, "language": "en", - "abstract": "Abstract:As the size of datasets continues to grow, machine learning applications are becoming increasingly limited by the amount of available computational power. Taking advantage of modern hardware requires using multiple parallel programming models targeted at different devices (e.g. CPUs and GPUs). However, programming these devices to run efficiently and correctly is difficult, error-prone, and results in software that is harder to read and maintain. We present OptiML, a domain-specific language (DSL) for machine learning. OptiML is an implicitly parallel, expressive and high perfromance alternative to MATLAB and C++. OptiML perfroms domain-specific analyses and optimizations and automatically generates CUDA code for GPUs. We show that OptiML outperfroms explicitly parallelized MATLAB code in nearly all cases.\n" + "abstract": "Abstract:As the size of datasets continues to grow, machine learning applications are becoming increasingly limited by the amount of available computational power. Taking advantage of modern hardware requires using multiple parallel programming models targeted at different devices (e.g. CPUs and GPUs). However, programming these devices to run efficiently and correctly is difficult, error-prone, and results in software that is harder to read and maintain. We present OptiML, a domain-specific language (DSL) for machine learning. OptiML is an implicitly parallel, expressive and high performance alternative to MATLAB and C++. OptiML performs domain-specific analyses and optimizations and automatically generates CUDA code for GPUs. We show that OptiML outperforms explicitly parallelized MATLAB code in nearly all cases." }, { "author": [ @@ -1215,14 +1215,14 @@ "Xing, Eric" ], "paper_title": "Infinite SVM: a Dirichlet Process Mixture of Large-margin Kernel Machines ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "EURASIP Workshop", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 617, + "page_form": 617, "page_to": 624, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present Infinite SVM (iSVM), a Dirichlet process mixture of large-margin kernel machines for multi-way classification. An iSVM enjoys the advantages of both Bayesian nonparametrics in handling the unknown number of mixing components, and large-margin kernel machines in robustly capturing local nonlinearity of complex data. We develop an efficient variational learning algorithm for posterior inference of iSVM, and we demonstrate the advantages of iSVM over Dirichlet process mixture of generalized linear models and other benchmarks on both synthetic and real Flickr image classification datasets.\n" + "abstract": "Abstract:We present Infinite SVM (iSVM), a Dirichlet process mixture of large-margin kernel machines for multi-way classification. An iSVM enjoys the advantages of both Bayesian nonparametrics in handling the unknown number of mixing components, and large-margin kernel machines in robustly capturing local nonlinearity of complex data. We develop an efficient variational learning algorithm for posterior inference of iSVM, and we demonstrate the advantages of iSVM over Dirichlet process mixture of generalized linear models and other benchmarks on both synthetic and real Flickr image classification datasets." }, { "author": [ @@ -1232,14 +1232,14 @@ "Carin, Lawrence" ], "paper_title": "On the Integration of Topic Modeling and Dictionary Learning ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Sequence Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 625, + "page_form": 625, "page_to": 632, "totle_page": 8, "language": "en", - "abstract": "Abstract:A new nonparametric Bayesian model is developed to integrate dictionary learning and topic model into a unified framework. The model is employed to analyze partially annotated images, with the dictionary learning perfromed directly on image patches. Efficient inference is perfromed with a Gibbs-slice sampler, and encouraging results are reported on widely used datasets.\n" + "abstract": "Abstract:A new nonparametric Bayesian model is developed to integrate dictionary learning and topic model into a unified framework. The model is employed to analyze partially annotated images, with the dictionary learning performed directly on image patches. Efficient inference is performed with a Gibbs-slice sampler, and encouraging results are reported on widely used datasets." }, { "author": [ @@ -1248,14 +1248,14 @@ "Kevin Murphy, University of British Columbia" ], "paper_title": "Piecewise Bounds for Estimating Bernoulli-Logistic Latent Gaussian Models ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "ROC Analysis in Artificial Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 633, + "page_form": 633, "page_to": 640, "totle_page": 8, "language": "en", - "abstract": "Abstract:Bernoulli-logistic latent Gaussian models (bLGMs) are a useful model class, but accurate parameter estimation is complicated by the fact that the marginal likelihood contains an intractable logistic-Gaussian integral. In this work, we propose the use of fixed piecewise linear and quadratic upper bounds to the logistic-log-partition (LLP) function as a way of circumventing this intractable integral. We describe a framework for approximately computing minimax optimal piecewise quadratic bounds, as well a generalized expectation maximization algorithm based on using piecewise bounds to estimate bLGMs. We prove a theoretical result relating the maximum error in the LLP bound to the maximum error in the marginal likelihood estimate. Finally, we present empirical results showing that piecewise bounds can be significantly more accurate than previously proposed variational bounds.\n" + "abstract": "Abstract:Bernoulli-logistic latent Gaussian models (bLGMs) are a useful model class, but accurate parameter estimation is complicated by the fact that the marginal likelihood contains an intractable logistic-Gaussian integral. In this work, we propose the use of fixed piecewise linear and quadratic upper bounds to the logistic-log-partition (LLP) function as a way of circumventing this intractable integral. We describe a framework for approximately computing minimax optimal piecewise quadratic bounds, as well a generalized expectation maximization algorithm based on using piecewise bounds to estimate bLGMs. We prove a theoretical result relating the maximum error in the LLP bound to the maximum error in the marginal likelihood estimate. Finally, we present empirical results showing that piecewise bounds can be significantly more accurate than previously proposed variational bounds." }, { "author": [ @@ -1264,14 +1264,14 @@ "Ben-David, Shai" ], "paper_title": "Access to Unlabeled Data can Speed up Prediction Time", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Emerging Trends in Engineering & Technology", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 641, + "page_form": 641, "page_to": 648, "totle_page": 8, "language": "en", - "abstract": "Abstract:Semi-supervised learning (SSL) addresses the problem of training a classifier using a small number of labeled examples and many unlabeled examples. Most previous work on SSL focused on how availability of unlabeled data can improve the accuracy of the learned classifiers. In this work we study how unlabeled data can be beneficial for constructing faster classifiers. We propose an SSL algorithmic framework which can utilize unlabeled examples for learning classifiers from a predefined set of fast classifiers. We fromally analyze conditions under which our algorithmic paradigm obtains significant improvements by the use of unlabeled data. As a side benefit of our analysis we propose a novel quantitative measure of the so-called cluster assumption. We demonstrate the potential merits of our approach by conducting experiments on the MNIST data set, showing that, when a sufficiently large unlabeled sample is available, a fast classifier can be learned from much fewer labeled examples than without such a sample.\n" + "abstract": "Abstract:Semi-supervised learning (SSL) addresses the problem of training a classifier using a small number of labeled examples and many unlabeled examples. Most previous work on SSL focused on how availability of unlabeled data can improve the accuracy of the learned classifiers. In this work we study how unlabeled data can be beneficial for constructing faster classifiers. We propose an SSL algorithmic framework which can utilize unlabeled examples for learning classifiers from a predefined set of fast classifiers. We formally analyze conditions under which our algorithmic paradigm obtains significant improvements by the use of unlabeled data. As a side benefit of our analysis we propose a novel quantitative measure of the so-called cluster assumption. We demonstrate the potential merits of our approach by conducting experiments on the MNIST data set, showing that, when a sufficiently large unlabeled sample is available, a fast classifier can be learned from much fewer labeled examples than without such a sample." }, { "author": [ @@ -1280,14 +1280,14 @@ "Marchand, Mario" ], "paper_title": "From PAC-Bayes Bounds to Quadratic Programs for Majority Votes ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Machine Learning; Models, Technologies and Applications", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 649, + "page_form": 649, "page_to": 656, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose to construct a weighted majority vote on a set of basis functions by minimizing a risk bound (called the C-bound) that depends on the first two moments of the margin of the Q-convex combination realized on the training data. This bound minimization algorithm turns out to be a quadratic program that can be efficiently solved. A first version of the algorithm is designed for the supervised inductive setting and turns out to be competitive with AdaBoost, MDBoost and the SVM. The second version of the algorithm, designed for the transductive setting, competes well with TSVM. We also propose a new PAC-Bayes theorem that bounds the difference between the \"true\" value of the C-bound and its empirical estimate and that, unexpectedly, contains no KL-divergence.\n" + "abstract": "Abstract:We propose to construct a weighted majority vote on a set of basis functions by minimizing a risk bound (called the C-bound) that depends on the first two moments of the margin of the Q-convex combination realized on the training data. This bound minimization algorithm turns out to be a quadratic program that can be efficiently solved. A first version of the algorithm is designed for the supervised inductive setting and turns out to be competitive with AdaBoost, MDBoost and the SVM. The second version of the algorithm, designed for the transductive setting, competes well with TSVM. We also propose a new PAC-Bayes theorem that bounds the difference between the \"true\" value of the C-bound and its empirical estimate and that, unexpectedly, contains no KL-divergence." }, { "author": [ @@ -1295,15 +1295,15 @@ "Hernandez-Orallo, Jose", "Ferri, C\\`{e}sar" ], - "paper_title": "A Coherent Interpretation of AUC as a Measure of Aggregated Classification Perfromance ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "paper_title": "A Coherent Interpretation of AUC as a Measure of Aggregated Classification Performance ", + "booktitle": "Recent Issues in Pattern Analysis and Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 657, + "page_form": 657, "page_to": 664, "totle_page": 8, "language": "en", - "abstract": "Abstract:The area under the ROC curve (AUC), a well-known measure of ranking perfromance, is also often used as a measure of classification perfromance, aggregating over decision thresholds as well as class and cost skews. However, David Hand has recently argued that AUC is fundamentally incoherent as a measure of aggregated classifier perfromance and proposed an alternative measure. Specifically, Hand derives a linear relationship between AUC and expected minimum loss, where the expectation is taken over a distribution of the misclassification cost parameter that depends on the model under consideration. Replacing this distribution with a Beta(2;2) distribution, Hand derives his alternative measure H. In this paper we offer an alternative, coherent interpretation of AUC as linearly related to expected loss. We use a distribution over cost parameter and a distribution over data points, both unifrom and hence model independent. Should one wish to consider only optimal thresholds, we demonstrate that a simple and more intuitive alternative to Hand’s H measure is already available in the from of the area under the cost curve.\n" + "abstract": "Abstract:The area under the ROC curve (AUC), a well-known measure of ranking performance, is also often used as a measure of classification performance, aggregating over decision thresholds as well as class and cost skews. However, David Hand has recently argued that AUC is fundamentally incoherent as a measure of aggregated classifier performance and proposed an alternative measure. Specifically, Hand derives a linear relationship between AUC and expected minimum loss, where the expectation is taken over a distribution of the misclassification cost parameter that depends on the model under consideration. Replacing this distribution with a Beta(2;2) distribution, Hand derives his alternative measure H. In this paper we offer an alternative, coherent interpretation of AUC as linearly related to expected loss. We use a distribution over cost parameter and a distribution over data points, both uniform and hence model independent. Should one wish to consider only optimal thresholds, we demonstrate that a simple and more intuitive alternative to Hand’s H measure is already available in the form of the area under the cost curve." }, { "author": [ @@ -1312,14 +1312,14 @@ "Sch\\\"{o}lkopf, Bernhard" ], "paper_title": "Support Vector Machines as Probabilistic Models ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Chinese Conference on Biometric Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 665, + "page_form": 665, "page_to": 672, "totle_page": 8, "language": "en", - "abstract": "Abstract:We show how the SVM can be viewed as a maximum likelihood estimate of a class of probabilistic models. This model class can be viewed as a reparametrization of the SVM in a similar vein to the $\\nu$-SVM reparametrizing the original SVM. It is not discriminative, but has a non-unifrom marginal. We illustrate the benefits of this new view by re-deriving and re-investigating two established SVM-related algorithms.\n" + "abstract": "Abstract:We show how the SVM can be viewed as a maximum likelihood estimate of a class of probabilistic models. This model class can be viewed as a reparametrization of the SVM in a similar vein to the $\\nu$-SVM reparametrizing the original SVM. It is not discriminative, but has a non-uniform marginal. We illustrate the benefits of this new view by re-deriving and re-investigating two established SVM-related algorithms." }, { "author": [ @@ -1330,14 +1330,14 @@ "Kalai, Adam" ], "paper_title": "Adaptively Learning the Crowd Kernel ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Asian Fuzzy System Society", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 673, + "page_form": 673, "page_to": 680, "totle_page": 8, "language": "en", - "abstract": "Abstract:We introduce an algorithm that, given n objects, learns a similarity matrix over all n^2 pairs, from crowdsourced data *alone*. The algorithm samples responses to adaptively chosen triplet-based relative-similarity queries. Each query has the from \"is object a more similar to b or to c?\" and is chosen to be maximally infromative given the preceding responses. The output is an embedding of the objects into Euclidean space (like MDS); we refer to this as the \"crowd kernel.\" SVMs reveal that the crowd kernel captures prominent and subtle features across a number of domains, such as \"is striped\" among neckties and \"vowel vs. consonant\" among letters.\n" + "abstract": "Abstract:We introduce an algorithm that, given n objects, learns a similarity matrix over all n^2 pairs, from crowdsourced data *alone*. The algorithm samples responses to adaptively chosen triplet-based relative-similarity queries. Each query has the form \"is object a more similar to b or to c?\" and is chosen to be maximally informative given the preceding responses. The output is an embedding of the objects into Euclidean space (like MDS); we refer to this as the \"crowd kernel.\" SVMs reveal that the crowd kernel captures prominent and subtle features across a number of domains, such as \"is striped\" among neckties and \"vowel vs. consonant\" among letters." }, { "author": [ @@ -1345,14 +1345,14 @@ "Teh, Yee Whye" ], "paper_title": "Bayesian Learning via Stochastic Gradient Langevin Dynamics", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Workshop on Attention and Performance in Computational Vision", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 681, + "page_form": 681, "page_to": 688, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper we propose a new framework for learning from large scale datasets based on iterative learning from small mini-batches. By adding the right amount of noise to a standard stochastic gradient optimization algorithm we show that the iterates will converge to samples from the true posterior distribution as we anneal the stepsize. This seamless transition between optimization and Bayesian posterior sampling provides an in-built protection against overfitting. We also propose a practical method for Monte Carlo estimates of posterior statistics which monitors a ``sampling threshold'' and collects samples after it has been surpassed. We apply the method to three models: a mixture of Gaussians, logistic regression and ICA with natural gradients.\n" + "abstract": "Abstract:In this paper we propose a new framework for learning from large scale datasets based on iterative learning from small mini-batches. By adding the right amount of noise to a standard stochastic gradient optimization algorithm we show that the iterates will converge to samples from the true posterior distribution as we anneal the stepsize. This seamless transition between optimization and Bayesian posterior sampling provides an in-built protection against overfitting. We also propose a practical method for Monte Carlo estimates of posterior statistics which monitors a ``sampling threshold'' and collects samples after it has been surpassed. We apply the method to three models: a mixture of Gaussians, logistic regression and ICA with natural gradients." }, { "author": [ @@ -1364,14 +1364,14 @@ "Ng, Andrew" ], "paper_title": "Multimodal Deep Learning", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Human Interactive Proofs", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 689, + "page_form": 689, "page_to": 696, "totle_page": 8, "language": "en", - "abstract": "Abstract:Deep networks have been successfully applied to unsupervised feature learning for single modalities (e.g., text, images or audio). In this work, we propose a novel application of deep networks to learn features over multiple modalities. We present a series of tasks for multimodal learning and show how to train deep networks that learn features to address these tasks. In particular, we demonstrate cross modality feature learning, where better features for one modality (e.g., video) can be learned if multiple modalities (e.g., audio and video) are present at feature learning time. Furthermore, we show how to learn a shared representation between modalities and evaluate it on a unique task, where the classifier is trained with audio-only data but tested with video-only data and vice-versa. Our methods are validated on the CUAVE and AVLetters datasets with an audio-visual speech classification task, demonstrating best published visual speech classification on AVLetters and effective shared representation learning.\n" + "abstract": "Abstract:Deep networks have been successfully applied to unsupervised feature learning for single modalities (e.g., text, images or audio). In this work, we propose a novel application of deep networks to learn features over multiple modalities. We present a series of tasks for multimodal learning and show how to train deep networks that learn features to address these tasks. In particular, we demonstrate cross modality feature learning, where better features for one modality (e.g., video) can be learned if multiple modalities (e.g., audio and video) are present at feature learning time. Furthermore, we show how to learn a shared representation between modalities and evaluate it on a unique task, where the classifier is trained with audio-only data but tested with video-only data and vice-versa. Our methods are validated on the CUAVE and AVLetters datasets with an audio-visual speech classification task, demonstrating best published visual speech classification on AVLetters and effective shared representation learning." }, { "author": [ @@ -1379,14 +1379,14 @@ "Scott, Clayton" ], "paper_title": "On the Robustness of Kernel Density M-Estimators ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Italian Workshop on Neural Neural Nets", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 697, + "page_form": 697, "page_to": 704, "totle_page": 8, "language": "en", - "abstract": "Abstract:We analyze a method for nonparametric density estimation that exhibits robustness to contamination of the training sample. This method achieves robustness by combining a traditional kernel density estimator (KDE) with ideas from classical M-estimation. The KDE based on a Gaussian kernel is interpreted as a sample mean in the associated reproducing kernel Hilbert space (RKHS). This mean is estimated robustly through the use of a robust loss, yielding the so-called robust kernel density estimator (RKDE). This robust sample mean can be found via a kernelized iteratively re-weighted least squares (IRWLS) algorithm. Our contributions are summarized as follows. First, we present a representer theorem for the RKDE, which gives an insight into the robustness of the RKDE. Second, we provide necessary and sufficient conditions for kernel IRWLS to converge to the global minimizer, in the Gaussian RKHS, of the objective function defining the RKDE. Third, characterize and provide a method for computing the influence function associated with the RKDE. Fourth, we illustrate the robustness of the RKDE through experiments on several data sets.\n" + "abstract": "Abstract:We analyze a method for nonparametric density estimation that exhibits robustness to contamination of the training sample. This method achieves robustness by combining a traditional kernel density estimator (KDE) with ideas from classical M-estimation. The KDE based on a Gaussian kernel is interpreted as a sample mean in the associated reproducing kernel Hilbert space (RKHS). This mean is estimated robustly through the use of a robust loss, yielding the so-called robust kernel density estimator (RKDE). This robust sample mean can be found via a kernelized iteratively re-weighted least squares (IRWLS) algorithm. Our contributions are summarized as follows. First, we present a representer theorem for the RKDE, which gives an insight into the robustness of the RKDE. Second, we provide necessary and sufficient conditions for kernel IRWLS to converge to the global minimizer, in the Gaussian RKHS, of the objective function defining the RKDE. Third, characterize and provide a method for computing the influence function associated with the RKDE. Fourth, we illustrate the robustness of the RKDE through experiments on several data sets." }, { "author": [ @@ -1394,14 +1394,14 @@ "III, Hal Daume" ], "paper_title": "Beam Search based MAP Estimates for the Indian Buffet Process ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Computing: Theory and Applications", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 705, + "page_form": 705, "page_to": 712, "totle_page": 8, "language": "en", - "abstract": "Abstract:Nonparametric latent feature models offer a flexible way to discover the latent features underlying the data, without having to a priori specify their number. The Indian Buffet Process (IBP) is a popular example of such a model. Inference in IBP based models, however, remains a challenge. Sampling techniques such as MCMC can be computationally expensive and can take a long time to converge to the stationary distribution. Variational techniques, although faster than sampling, can be difficult to design, and can still remain slow on large data. In many problems, however, we only seek a maximum a posteriori (MAP) estimate of the latent feature assignment matrix. For such cases, we show that techniques such as beam search can give fast, approximate MAP estimates in the IBP based models. If samples from the posterior are desired, these MAP estimates can also serve as sensible initializers for MCMC based algorithms. Experimental results on a variety of datasets suggest that our algorithms can be a computationally viable alternative to Gibbs sampling, the particle filter, and variational inference based approaches for the IBP, and also perfrom better than other heuristics such as greedy search.\n" + "abstract": "Abstract:Nonparametric latent feature models offer a flexible way to discover the latent features underlying the data, without having to a priori specify their number. The Indian Buffet Process (IBP) is a popular example of such a model. Inference in IBP based models, however, remains a challenge. Sampling techniques such as MCMC can be computationally expensive and can take a long time to converge to the stationary distribution. Variational techniques, although faster than sampling, can be difficult to design, and can still remain slow on large data. In many problems, however, we only seek a maximum a posteriori (MAP) estimate of the latent feature assignment matrix. For such cases, we show that techniques such as beam search can give fast, approximate MAP estimates in the IBP based models. If samples from the posterior are desired, these MAP estimates can also serve as sensible initializers for MCMC based algorithms. Experimental results on a variety of datasets suggest that our algorithms can be a computationally viable alternative to Gibbs sampling, the particle filter, and variational inference based approaches for the IBP, and also perform better than other heuristics such as greedy search." }, { "author": [ @@ -1411,14 +1411,14 @@ "Xiao, Lin" ], "paper_title": "Optimal Distributed Online Prediction ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Turkish Symposium on Artificial Intelligence and Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 713, + "page_form": 713, "page_to": 720, "totle_page": 8, "language": "en", - "abstract": "Abstract:Online prediction methods are typically studied as serial algorithms running on a single processor. In this paper, we present the distributed mini-batch (DMB) framework, a method of converting a serial gradient-based online algorithm into a distributed algorithm, and prove an asymptotically optimal regret bound for smooth convex loss functions and stochastic examples. Our analysis explicitly takes into account communication latencies between computing nodes in a network. We also present robust variants, which are resilient to failures and node heterogeneity in an synchronous distributed environment. Our method can also be used for distributed stochastic optimization, attaining an asymptotically linear speedup. Finally, we empirically demonstrate the merits of our approach on large-scale online prediction problems.\n" + "abstract": "Abstract:Online prediction methods are typically studied as serial algorithms running on a single processor. In this paper, we present the distributed mini-batch (DMB) framework, a method of converting a serial gradient-based online algorithm into a distributed algorithm, and prove an asymptotically optimal regret bound for smooth convex loss functions and stochastic examples. Our analysis explicitly takes into account communication latencies between computing nodes in a network. We also present robust variants, which are resilient to failures and node heterogeneity in an synchronous distributed environment. Our method can also be used for distributed stochastic optimization, attaining an asymptotically linear speedup. Finally, we empirically demonstrate the merits of our approach on large-scale online prediction problems." }, { "author": [ @@ -1427,14 +1427,14 @@ "Ghahramani, Zoubin" ], "paper_title": "Message Passing Algorithms for the Dirichlet Diffusion Tree ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Biometric Recognition Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 721, + "page_form": 721, "page_to": 728, "totle_page": 8, "language": "en", - "abstract": "Abstract:We demonstrate efficient approximate inference for the Dirichlet Diffusion Tree, a Bayesian nonparametric prior over tree structures. Although DDTs provide a powerful and elegant approach for modeling hierarchies they haven't seen much use to date. One problem is the computational cost of MCMC inference. We provide the first deterministic approximate inference methods for DDT models and show excellent perfromance compared to the MCMC alternative. We present message passing algorithms to approximate the Bayesian model evidence for a specific tree. This is used to drive sequential tree building and greedy search to find optimal tree structures, corresponding to hierarchical clusterings of the data. We demonstrate appropriate observation models for continuous and binary data. The empirical perfromance of our method is very close to the computationally expensive MCMC alternative on a density estimation problem, and significantly outperfroms kernel density estimators.\n" + "abstract": "Abstract:We demonstrate efficient approximate inference for the Dirichlet Diffusion Tree, a Bayesian nonparametric prior over tree structures. Although DDTs provide a powerful and elegant approach for modeling hierarchies they haven't seen much use to date. One problem is the computational cost of MCMC inference. We provide the first deterministic approximate inference methods for DDT models and show excellent performance compared to the MCMC alternative. We present message passing algorithms to approximate the Bayesian model evidence for a specific tree. This is used to drive sequential tree building and greedy search to find optimal tree structures, corresponding to hierarchical clusterings of the data. We demonstrate appropriate observation models for continuous and binary data. The empirical performance of our method is very close to the computationally expensive MCMC alternative on a density estimation problem, and significantly outperforms kernel density estimators." }, { "author": [ @@ -1444,14 +1444,14 @@ "Urtasun, Raquel" ], "paper_title": "Convex Max-Product over Compact Sets for Protein Folding", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "MirrorBot Project", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 729, + "page_form": 729, "page_to": 736, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper we present an approach to inference in graphical models with mixture of discrete and bounded continuous variables. In particular, we extend convex max-product to deal with these hybrid models and derive the conditions under which our approach is guaranteed to produce the MAP assignment. When dealing with continuous variables the messages are functions. We investigate a multi-grid approach which can be viewed as a piecewise constant representation of these messages. While this approach provides the theoretical guarantees it is not very practical. Inspired by this, we further propose a particle convex max-product algorithm that significantly outperfroms existing particle methods in the task of protein folding and perfroms comparable to the state-of-the art while using a smaller amount of prior knowledge.\n" + "abstract": "Abstract:In this paper we present an approach to inference in graphical models with mixture of discrete and bounded continuous variables. In particular, we extend convex max-product to deal with these hybrid models and derive the conditions under which our approach is guaranteed to produce the MAP assignment. When dealing with continuous variables the messages are functions. We investigate a multi-grid approach which can be viewed as a piecewise constant representation of these messages. While this approach provides the theoretical guarantees it is not very practical. Inspired by this, we further propose a particle convex max-product algorithm that significantly outperforms existing particle methods in the task of protein folding and performs comparable to the state-of-the art while using a smaller amount of prior knowledge." }, { "author": [ @@ -1459,14 +1459,14 @@ "Stone, Peter" ], "paper_title": "Structure Learning in Ergodic Factored MDPs without Knowledge of the Transition Function's In-Degree ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Deterministic and Statistical Methods in Machine Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 737, + "page_form": 737, "page_to": 744, "totle_page": 8, "language": "en", - "abstract": "Abstract:This paper introduces Learn Structure and Exploit RMax (LSE-RMax), a novel model based structure learning algorithm for ergodic factored-state MDPs. Given a planning horizon that satisfies a condition, LSE-RMax provably guarantees a return very close to the optimal return, with a high certainty, without requiring any prior knowledge of the in-degree of the transition function as input. LSE-RMax is fully implemented with a thorough analysis of its sample complexity. We also present empirical results demonstrating its effectiveness compared to prior approaches to the problem.\n" + "abstract": "Abstract:This paper introduces Learn Structure and Exploit RMax (LSE-RMax), a novel model based structure learning algorithm for ergodic factored-state MDPs. Given a planning horizon that satisfies a condition, LSE-RMax provably guarantees a return very close to the optimal return, with a high certainty, without requiring any prior knowledge of the in-degree of the transition function as input. LSE-RMax is fully implemented with a thorough analysis of its sample complexity. We also present empirical results demonstrating its effectiveness compared to prior approaches to the problem." }, { "author": [ @@ -1476,14 +1476,14 @@ "Joulin, Armand" ], "paper_title": "Clusterpath: an Algorithm for Clustering using Convex Fusion Penalties ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Brain, Vision, and Artificial Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 745, + "page_form": 745, "page_to": 752, "totle_page": 8, "language": "en", - "abstract": "Abstract: We present a new clustering algorithm by proposing a convex relaxation of hierarchical clustering, which results in a family of objective functions with a natural geometric interpretation. We give efficient algorithms for calculating the continuous regularization path of solutions, and discuss relative advantages of the parameters. Our method experimentally gives state-of-the-art results similar to spectral clustering for non-convex clusters, and has the added benefit of learning a tree structure from the data.\n" + "abstract": "Abstract: We present a new clustering algorithm by proposing a convex relaxation of hierarchical clustering, which results in a family of objective functions with a natural geometric interpretation. We give efficient algorithms for calculating the continuous regularization path of solutions, and discuss relative advantages of the parameters. Our method experimentally gives state-of-the-art results similar to spectral clustering for non-convex clusters, and has the added benefit of learning a tree structure from the data." }, { "author": [ @@ -1492,14 +1492,14 @@ "Airoldi, Edo" ], "paper_title": "Tree preserving embedding ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Neural Networks and Computational Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 753, + "page_form": 753, "page_to": 760, "totle_page": 8, "language": "en", - "abstract": "Abstract:Visualization techniques for complex data are a workhorse of modern scientific pursuits. The goal of visualization is to embed high dimensional data in a low dimensional space, while preserving structure in the data relevant to exploratory data analysis, such as the existence of clusters. However, existing visualization methods often either entirely fail to preserve clusters in embeddings due to the crowding problem or can only preserve clusters at a single resolution. Here, we develop a new approach to visualization, tree preserving embedding (TPE). Our approach takes advantage of the topological notion of connectedness to provably preserve clusters at all resolutions. Our perfromance guarantee holds for finite samples, which makes TPE a robust method for applications. Our approach suggests new strategies for robust data visualization in practice.\n" + "abstract": "Abstract:Visualization techniques for complex data are a workhorse of modern scientific pursuits. The goal of visualization is to embed high dimensional data in a low dimensional space, while preserving structure in the data relevant to exploratory data analysis, such as the existence of clusters. However, existing visualization methods often either entirely fail to preserve clusters in embeddings due to the crowding problem or can only preserve clusters at a single resolution. Here, we develop a new approach to visualization, tree preserving embedding (TPE). Our approach takes advantage of the topological notion of connectedness to provably preserve clusters at all resolutions. Our performance guarantee holds for finite samples, which makes TPE a robust method for applications. Our approach suggests new strategies for robust data visualization in practice." }, { "author": [ @@ -1509,14 +1509,14 @@ "Maryam Fazel" ], "paper_title": "Clustering by Left-Stochastic Matrix Factorization ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Artificial Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 761, + "page_form": 761, "page_to": 768, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose clustering samples given their pairwise similarities by factorizing the similarity matrix into the product of a cluster probability matrix and its transpose. We propose a rotation-based algorithm to compute this left-stochastic decomposition (LSD). Theoretical results link the LSD clustering method to a soft kernel k-means clustering, give conditions for when the factorization and clustering are unique, and provide error bounds. Experimental results on simulated and real similarity datasets show that the proposed method reliably provides accurate clusterings.\n" + "abstract": "Abstract:We propose clustering samples given their pairwise similarities by factorizing the similarity matrix into the product of a cluster probability matrix and its transpose. We propose a rotation-based algorithm to compute this left-stochastic decomposition (LSD). Theoretical results link the LSD clustering method to a soft kernel k-means clustering, give conditions for when the factorization and clustering are unique, and provide error bounds. Experimental results on simulated and real similarity datasets show that the proposed method reliably provides accurate clusterings." }, { "author": [ @@ -1525,14 +1525,14 @@ "Carin, Lawrence" ], "paper_title": "The Infinite Regionalized Policy Representation ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Large-Scale Knowledge Resources", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 769, + "page_form": 769, "page_to": 776, "totle_page": 8, "language": "en", - "abstract": "Abstract:We introduce the infinite regionalized policy presentation (iRPR), as a nonparametric policy for reinforcement learning in partially observable Markov decision processes (POMDPs). The iRPR assumes an unbounded set of decision states a priori, and infers the number of states to represent the policy given the experiences. We propose algorithms for learning the number of decision states while maintaining a proper balance between exploration and exploitation. Convergence analysis is provided, along with perfromance evaluations on benchmark problems.\n" + "abstract": "Abstract:We introduce the infinite regionalized policy presentation (iRPR), as a nonparametric policy for reinforcement learning in partially observable Markov decision processes (POMDPs). The iRPR assumes an unbounded set of decision states a priori, and infers the number of states to represent the policy given the experiences. We propose algorithms for learning the number of decision states while maintaining a proper balance between exploration and exploitation. Convergence analysis is provided, along with performance evaluations on benchmark problems." }, { "author": [ @@ -1543,14 +1543,14 @@ "McCallum, Andrew" ], "paper_title": "SampleRank: Training Factor Graphs with Atomic Gradients", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Knowledge Representation and Organization in Machine Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 777, + "page_form": 777, "page_to": 784, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present SampleRank, an alternative to contrastive divergence (CD) for estimating parameters in complex graphical models. SampleRank harnesses a user-provided loss function to distribute stochastic gradients across an MCMC chain. As a result, parameter updates can be computed between arbitrary MCMC states. SampleRank is not only faster than CD, but also achieves better accuracy in practice (up to 23\\% error reduction on noun-phrase coreference).\n" + "abstract": "Abstract:We present SampleRank, an alternative to contrastive divergence (CD) for estimating parameters in complex graphical models. SampleRank harnesses a user-provided loss function to distribute stochastic gradients across an MCMC chain. As a result, parameter updates can be computed between arbitrary MCMC states. SampleRank is not only faster than CD, but also achieves better accuracy in practice (up to 23\\% error reduction on noun-phrase coreference)." }, { "author": [ @@ -1559,14 +1559,14 @@ "Carin, Lawrence" ], "paper_title": "Tree-Structured Infinite Sparse Factor Model", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "European Summer School on Multi-Agent Control", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 785, + "page_form": 785, "page_to": 792, "totle_page": 8, "language": "en", - "abstract": "Abstract:A new tree-structured multiplicative gamma process (TMGP) is developed, for inferring the depth of a tree-based factor-analysis model. This new model is coupled with the nested Chinese restaurant process, to nonparametrically infer the depth and width (structure) of the tree. In addition to developing the model, theoretical properties of the TMGP are addressed, and a novel MCMC sampler is developed. The structure of the inferred tree is used to learn relationships between high-dimensional data, and the model is also applied to compressive sensing and interpolation of incomplete images.\n" + "abstract": "Abstract:A new tree-structured multiplicative gamma process (TMGP) is developed, for inferring the depth of a tree-based factor-analysis model. This new model is coupled with the nested Chinese restaurant process, to nonparametrically infer the depth and width (structure) of the tree. In addition to developing the model, theoretical properties of the TMGP are addressed, and a novel MCMC sampler is developed. The structure of the inferred tree is used to learn relationships between high-dimensional data, and the model is also applied to compressive sensing and interpolation of incomplete images." }, { "author": [ @@ -1575,14 +1575,14 @@ "Gurevich, Maxim" ], "paper_title": "Preserving Personalized Pagerank in Subgraphs", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Subspace, Latent Structure and Feature Selection", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 793, + "page_form": 793, "page_to": 800, "totle_page": 8, "language": "en", - "abstract": "Abstract:Choosing a subgraph that can concisely represent a large real-world graph is useful in many scenarios. The usual strategy employed is to sample nodes so that the induced subgraph matches the original graph’s degree distribution, clustering coefficient, etc., but no attempt is made to preserve fine-grained relationships between nodes, which are vital for applications like clustering, classification, and ranking. In this work, we model such relationships via the notion of Personalized PageRank Value (PPV). We show that induced subgraphs output by current sampling methods do not preserve PPVs, and propose algorithms for creating PPV-preserving subgraphs on any given subset of graph nodes. Experiments on three large real-world graphs show that the subgraphs created by our method improve upon induced subgraphs in terms of preserving PPVs, clustering accuracy, and maintaining basic graph properties.\n" + "abstract": "Abstract:Choosing a subgraph that can concisely represent a large real-world graph is useful in many scenarios. The usual strategy employed is to sample nodes so that the induced subgraph matches the original graph’s degree distribution, clustering coefficient, etc., but no attempt is made to preserve fine-grained relationships between nodes, which are vital for applications like clustering, classification, and ranking. In this work, we model such relationships via the notion of Personalized PageRank Value (PPV). We show that induced subgraphs output by current sampling methods do not preserve PPVs, and propose algorithms for creating PPV-preserving subgraphs on any given subset of graph nodes. Experiments on three large real-world graphs show that the subgraphs created by our method improve upon induced subgraphs in terms of preserving PPVs, clustering accuracy, and maintaining basic graph properties." }, { "author": [ @@ -1591,14 +1591,14 @@ "Wu, Mingrui" ], "paper_title": "Hierarchical Classification via Orthogonal Transfer ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Computer Recognition Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 801, + "page_form": 801, "page_to": 808, "totle_page": 8, "language": "en", - "abstract": "Abstract:We consider multiclass classification problems where the set of labels are organized hierarchically as a category tree. We associate each node in the tree with a classifier and classify the examples recursively from the root to the leaves. We propose a hierarchical Support Vector Machine (SVM) that encourages the classifier at each node to be different from the classifiers at its ancestors. More specifically, we introduce regularizations that force the normal vector of the classifying hyperplane at each node to be orthogonal to those at its ancestors as much as possible. We establish conditions under which training such a hierarchical SVM is a convex optimization problem, and develop an efficient dual-averaging method for solving it. We evaluate the method on a number of real-world text categorization tasks and obtain state-of-the-art perfromance.\n" + "abstract": "Abstract:We consider multiclass classification problems where the set of labels are organized hierarchically as a category tree. We associate each node in the tree with a classifier and classify the examples recursively from the root to the leaves. We propose a hierarchical Support Vector Machine (SVM) that encourages the classifier at each node to be different from the classifiers at its ancestors. More specifically, we introduce regularizations that force the normal vector of the classifying hyperplane at each node to be orthogonal to those at its ancestors as much as possible. We establish conditions under which training such a hierarchical SVM is a convex optimization problem, and develop an efficient dual-averaging method for solving it. We evaluate the method on a number of real-world text categorization tasks and obtain state-of-the-art performance." }, { "author": [ @@ -1607,28 +1607,28 @@ "Kriegel, Hans-Peter" ], "paper_title": "A Three-Way Model for Collective Learning on Multi-Relational Data", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Nonlinear Analyses and Algorithms for Speech Processing", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 809, + "page_form": 809, "page_to": 816, "totle_page": 8, "language": "en", - "abstract": "Abstract:Relational learning is becoming increasingly important in many areas of application. Here, we present a novel approach to relational learning based on the factorization of a three-way tensor. We show that unlike other tensor approaches, our method is able to perfrom collective learning via the latent components of the model and provide an efficient algorithm to compute the factorization. We substantiate our theoretical considerations regarding the collective learning capabilities of our model by the means of experiments on both a new dataset and a dataset commonly used in entity resolution. Furthermore, we show on common benchmark datasets that our approach achieves better or on-par results, if compared to current state-of-the-art relational learning solutions, while it is significantly faster to compute.\n" + "abstract": "Abstract:Relational learning is becoming increasingly important in many areas of application. Here, we present a novel approach to relational learning based on the factorization of a three-way tensor. We show that unlike other tensor approaches, our method is able to perform collective learning via the latent components of the model and provide an efficient algorithm to compute the factorization. We substantiate our theoretical considerations regarding the collective learning capabilities of our model by the means of experiments on both a new dataset and a dataset commonly used in entity resolution. Furthermore, we show on common benchmark datasets that our approach achieves better or on-par results, if compared to current state-of-the-art relational learning solutions, while it is significantly faster to compute." }, { "author": [ "Neumann, Gerhard" ], "paper_title": "Variational Inference for Policy Search in changing situations ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Signal Processing, Pattern Recognition, and Applications", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 817, + "page_form": 817, "page_to": 824, "totle_page": 8, "language": "en", - "abstract": "Abstract:Many policy search algorithms minimize the Kullback-Leibler (KL) divergence to a certain target distribution in order to fit their policy. The commonly used KL-divergence forces the resulting policy to be 'reward-attracted'. The policy tries to reproduce all positively rewarded experience while negative experience is neglected. However, the KL-divergence is not symmetric and we can also minimize the the reversed KL-divergence, which is typically used in variational inference. The policy now becomes 'cost-averse'. It tries to avoid reproducing any negatively-rewarded experience while maximizing exploration. Due to this 'cost-averseness' of the policy, Variational Inference for Policy Search (VIP) has several interesting properties. It requires no kernel-bandwith nor exploration rate, such settings are determined automatically by the inference. The algorithm meets the perfromance of state-of-the-art methods while being applicable to simultaneously learning in multiple situations. We concentrate on using VIP for policy search in robotics. We apply our algorithm to learn dynamic counterbalancing of different kinds of pushes with a human-like 4-link robot.\n" + "abstract": "Abstract:Many policy search algorithms minimize the Kullback-Leibler (KL) divergence to a certain target distribution in order to fit their policy. The commonly used KL-divergence forces the resulting policy to be 'reward-attracted'. The policy tries to reproduce all positively rewarded experience while negative experience is neglected. However, the KL-divergence is not symmetric and we can also minimize the the reversed KL-divergence, which is typically used in variational inference. The policy now becomes 'cost-averse'. It tries to avoid reproducing any negatively-rewarded experience while maximizing exploration. Due to this 'cost-averseness' of the policy, Variational Inference for Policy Search (VIP) has several interesting properties. It requires no kernel-bandwith nor exploration rate, such settings are determined automatically by the inference. The algorithm meets the performance of state-of-the-art methods while being applicable to simultaneously learning in multiple situations. We concentrate on using VIP for policy search in robotics. We apply our algorithm to learn dynamic counterbalancing of different kinds of pushes with a human-like 4-link robot." }, { "author": [ @@ -1638,14 +1638,14 @@ "Usunier, Nicolas" ], "paper_title": "Learning Scoring Functions with Order-Preserving Losses and Standardized Supervision ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Workshop on Intelligent Computing in Pattern Analysis/Synthesis", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 825, + "page_form": 825, "page_to": 832, "totle_page": 8, "language": "en", - "abstract": "Abstract:We address the problem of designing surrogate losses for learning scoring functions in the context of label ranking. We extend to ranking problems a notion of order preserving losses previously introduced for multiclass classi?cation, and show that these losses lead to consistent fromulations with respect to a family of ranking evaluation metrics. An order-preserving loss can be tailored for a given evaluation metric by appropriately setting some weights depending on this metric and the observed supervision. These weights, called the standard from of the supervision, do not always exist, but we show that previous consistency results for ranking were proved in special cases where they do. We then evaluate a new pairwise loss consistent with the (Normalized) Discounted Cumulative Gain on benchmark datasets.\n" + "abstract": "Abstract:We address the problem of designing surrogate losses for learning scoring functions in the context of label ranking. We extend to ranking problems a notion of order preserving losses previously introduced for multiclass classi?cation, and show that these losses lead to consistent formulations with respect to a family of ranking evaluation metrics. An order-preserving loss can be tailored for a given evaluation metric by appropriately setting some weights depending on this metric and the observed supervision. These weights, called the standard form of the supervision, do not always exist, but we show that previous consistency results for ranking were proved in special cases where they do. We then evaluate a new pairwise loss consistent with the (Normalized) Discounted Cumulative Gain on benchmark datasets." }, { "author": [ @@ -1656,14 +1656,14 @@ "Bengio, Yoshua" ], "paper_title": "Contractive Auto-Encoders: Explicit Invariance During Feature Extraction ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Analysis of Neural Network Applications", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 833, + "page_form": 833, "page_to": 840, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present in this paper a novel approach for training deterministic auto-encoders. We show that by adding a well chosen penalty term to the classical reconstruction cost function, we can achieve results that equal or surpass those attained by other regularized auto-encoders as well as denoising auto-encoders on a range of datasets. This penalty term corresponds to the Frobenius norm of the Jacobian matrix of the encoder activations with respect to the input. We show that this penalty term results in a localized space contraction which in turn yields robust features on the activation layer. Furthermore, we show how this penalty term is related to both regularized auto-encoders and denoising auto-encoders and how it can be seen as a link between deterministic and non-deterministic auto-encoders. We find empirically that this penalty helps to carve a representation that better captures the local directions of variation dictated by the data, corresponding to a lower-dimensional non-linear manifold, while being more invariant to the vast majority of directions orthogonal to the manifold. Finally, we show that by using the learned features to initialize an MLP, we achieve state of the art classification error on a range of datasets, surpassing other methods of pre-training.\n" + "abstract": "Abstract:We present in this paper a novel approach for training deterministic auto-encoders. We show that by adding a well chosen penalty term to the classical reconstruction cost function, we can achieve results that equal or surpass those attained by other regularized auto-encoders as well as denoising auto-encoders on a range of datasets. This penalty term corresponds to the Frobenius norm of the Jacobian matrix of the encoder activations with respect to the input. We show that this penalty term results in a localized space contraction which in turn yields robust features on the activation layer. Furthermore, we show how this penalty term is related to both regularized auto-encoders and denoising auto-encoders and how it can be seen as a link between deterministic and non-deterministic auto-encoders. We find empirically that this penalty helps to carve a representation that better captures the local directions of variation dictated by the data, corresponding to a lower-dimensional non-linear manifold, while being more invariant to the vast majority of directions orthogonal to the manifold. Finally, we show that by using the learned features to initialize an MLP, we achieve state of the art classification error on a range of datasets, surpassing other methods of pre-training." }, { "author": [ @@ -1671,14 +1671,14 @@ "Titsias, Michalis" ], "paper_title": "Variational Heteroscedastic Gaussian Process Regression", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Machine Learning Summer School", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 841, + "page_form": 841, "page_to": 848, "totle_page": 8, "language": "en", - "abstract": "Abstract:Standard Gaussian processes (GPs) model observations' noise as constant throughout input space. This is often a too restrictive assumption, but one that is needed for GP inference to be tractable. In this work we present a non-standard variational approximation that allows accurate inference in heteroscedastic GPs (i.e., under input-dependent noise conditions). Computational cost is roughly twice that of the standard GP, and also scales as O(n^3). Accuracy is verified by comparing with the golden standard MCMC and its effectiveness is illustrated on several synthetic and real datasets of diverse characteristics. An application to volatility forecasting is also considered.\n" + "abstract": "Abstract:Standard Gaussian processes (GPs) model observations' noise as constant throughout input space. This is often a too restrictive assumption, but one that is needed for GP inference to be tractable. In this work we present a non-standard variational approximation that allows accurate inference in heteroscedastic GPs (i.e., under input-dependent noise conditions). Computational cost is roughly twice that of the standard GP, and also scales as O(n^3). Accuracy is verified by comparing with the golden standard MCMC and its effectiveness is illustrated on several synthetic and real datasets of diverse characteristics. An application to volatility forecasting is also considered." }, { "author": [ @@ -1686,14 +1686,14 @@ "Ihler, Alexander" ], "paper_title": "Bounding the Partition Function using Holder's Inequality ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Simulationstechnik", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 849, + "page_form": 849, "page_to": 856, "totle_page": 8, "language": "en", - "abstract": "Abstract:We describe an algorithm for approximate inference in graphical models based on Holder's inequality that provides upper and lower bounds on common summation problems such as computing the partition function or probability of evidence in a graphical model. Our algorithm unifies and extends several existing approaches, including variable elimination techniques such as mini-bucket elimination and variational methods such as tree reweighted belief propagation and conditional entropy decomposition. We show that our method inherits benefits from each approach to provide significantly better bounds on sum-product tasks.\n" + "abstract": "Abstract:We describe an algorithm for approximate inference in graphical models based on Holder's inequality that provides upper and lower bounds on common summation problems such as computing the partition function or probability of evidence in a graphical model. Our algorithm unifies and extends several existing approaches, including variable elimination techniques such as mini-bucket elimination and variational methods such as tree reweighted belief propagation and conditional entropy decomposition. We show that our method inherits benefits from each approach to provide significantly better bounds on sum-product tasks." }, { "author": [ @@ -1703,14 +1703,14 @@ "Smyth, Padhraic" ], "paper_title": "Dynamic Egocentric Models for Citation Networks", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Aachener Symposium für Signaltheorie", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 857, + "page_form": 857, "page_to": 864, "totle_page": 8, "language": "en", - "abstract": "Abstract:The analysis of the fromation and evolution of networks over time is of fundamental importance to social science, biology, and many other fields. While longitudinal network data sets are increasingly being recorded at the granularity of individual time-stamped events, most studies only focus on collapsed cross-sectional snapshots of the network. In this paper, we introduce a dynamic egocentric framework that models continuous-time network data using multivariate counting processes. For inference, an efficient partial likelihood approach is used, allowing our methods to scale to large networks. We apply our techniques to various citation networks and demonstrate the predictive power and interpretability of the learned statistical models.\n" + "abstract": "Abstract:The analysis of the formation and evolution of networks over time is of fundamental importance to social science, biology, and many other fields. While longitudinal network data sets are increasingly being recorded at the granularity of individual time-stamped events, most studies only focus on collapsed cross-sectional snapshots of the network. In this paper, we introduce a dynamic egocentric framework that models continuous-time network data using multivariate counting processes. For inference, an efficient partial likelihood approach is used, allowing our methods to scale to large networks. We apply our techniques to various citation networks and demonstrate the predictive power and interpretability of the learned statistical models." }, { "author": [ @@ -1720,14 +1720,14 @@ "Trikalinos, Thomas" ], "paper_title": "The Constrained Weight Space SVM: Learning with Ranked Features ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Tbilisi Symposium on Logic, Language, and Computation", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 865, + "page_form": 865, "page_to": 872, "totle_page": 8, "language": "en", - "abstract": "Abstract:Applying supervised learning methods to new classification tasks requires domain experts to label sufficient training data for the classifier to achieve acceptable perfromance. It is desirable to mitigate this annotation effort. To this end, a pertinent observation is that instance labels are often an indirect from of supervision; it may be more efficient to impart domain knowledge directly to the model in the from of labeled features. We present a novel algorithm for exploiting such domain knowledge which we call the Constrained Weight Space SVM (CW-SVM). In addition to exploiting binary labeled features, our approach allows domain experts to provide ranked features, and, more generally, to express arbitrary expected relationships between sets of features. Our empirical results show that the CW-SVM outperfroms both baseline supervised learning strategies and previously proposed methods for learning with labeled features.\n" + "abstract": "Abstract:Applying supervised learning methods to new classification tasks requires domain experts to label sufficient training data for the classifier to achieve acceptable performance. It is desirable to mitigate this annotation effort. To this end, a pertinent observation is that instance labels are often an indirect form of supervision; it may be more efficient to impart domain knowledge directly to the model in the form of labeled features. We present a novel algorithm for exploiting such domain knowledge which we call the Constrained Weight Space SVM (CW-SVM). In addition to exploiting binary labeled features, our approach allows domain experts to provide ranked features, and, more generally, to express arbitrary expected relationships between sets of features. Our empirical results show that the CW-SVM outperforms both baseline supervised learning strategies and previously proposed methods for learning with labeled features." }, { "author": [ @@ -1737,14 +1737,14 @@ "Sanghavi, Sujay" ], "paper_title": "Robust Matrix Completion and Corrupted Columns ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Workshop Farbbildverarbeitung", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 873, + "page_form": 873, "page_to": 880, "totle_page": 8, "language": "en", - "abstract": "Abstract:This paper considers the problem of matrix completion, when some number of the columns are arbitrarily corrupted. It is well-known that standard algorithms for matrix completion can return arbitrarily poor results, if even a single column is corrupted. What can be done if a large number, or even a constant fraction of columns are corrupted? In this paper, we study this very problem, and develop an robust and efficient algorithm for its solution. One direct application comes from robust collaborative filtering. Here, some number of users are so-called manipulators, and try to skew the predictions of the algorithm. Significantly, our results hold {\\it without any assumptions on the observed entries of the manipulated columns}.\n" + "abstract": "Abstract:This paper considers the problem of matrix completion, when some number of the columns are arbitrarily corrupted. It is well-known that standard algorithms for matrix completion can return arbitrarily poor results, if even a single column is corrupted. What can be done if a large number, or even a constant fraction of columns are corrupted? In this paper, we study this very problem, and develop an robust and efficient algorithm for its solution. One direct application comes from robust collaborative filtering. Here, some number of users are so-called manipulators, and try to skew the predictions of the algorithm. Significantly, our results hold {\\it without any assumptions on the observed entries of the manipulated columns}." }, { "author": [ @@ -1755,14 +1755,14 @@ "How, Jonathan" ], "paper_title": "Online Discovery of Feature Dependencies ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Automatic Learning and Real-Time", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 881, + "page_form": 881, "page_to": 888, "totle_page": 8, "language": "en", - "abstract": "Abstract:Online representational expansion techniques have improved the learning speed of existing reinforcement learning (RL) algorithms in low dimensional domains, yet existing online expansion methods do not scale well to high dimensional problems. We conjecture that one of the main difficulties limiting this scaling is that features defined over the full-dimensional state space often generalize poorly. Hence, we introduce incremental Feature Dependency Discovery (iFDD) as a computationally-inexpensive method for representational expansion that can be combined with any online, value-based RL method that uses binary features. Unlike other online expansion techniques, iFDD creates new features in low dimensional subspaces of the full state space where feedback errors persist. We provide convergence and computational complexity guarantees for iFDD, as well as showing empirically that iFDD scales well to high dimensional multi-agent planning domains with hundreds of millions of state-action pairs.\n" + "abstract": "Abstract:Online representational expansion techniques have improved the learning speed of existing reinforcement learning (RL) algorithms in low dimensional domains, yet existing online expansion methods do not scale well to high dimensional problems. We conjecture that one of the main difficulties limiting this scaling is that features defined over the full-dimensional state space often generalize poorly. Hence, we introduce incremental Feature Dependency Discovery (iFDD) as a computationally-inexpensive method for representational expansion that can be combined with any online, value-based RL method that uses binary features. Unlike other online expansion techniques, iFDD creates new features in low dimensional subspaces of the full state space where feedback errors persist. We provide convergence and computational complexity guarantees for iFDD, as well as showing empirically that iFDD scales well to high dimensional multi-agent planning domains with hundreds of millions of state-action pairs." }, { "author": [ @@ -1771,14 +1771,14 @@ "Blei, David" ], "paper_title": "Variational Inference for Stick-Breaking Beta Process Priors ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Artificial Neural Neutworks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 889, + "page_form": 889, "page_to": 896, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present a variational Bayesian inference algorithm for the stick-breaking construction of the beta process. We derive an alternate representation of the beta process that is amenable to variational inference, and present a bound relating the truncated beta process to its infinite counterpart. We assess perfromance on two matrix factorization problems, using a non-negative factorization model and a linear-Gaussian model.\n" + "abstract": "Abstract:We present a variational Bayesian inference algorithm for the stick-breaking construction of the beta process. We derive an alternate representation of the beta process that is amenable to variational inference, and present a bound relating the truncated beta process to its infinite counterpart. We assess performance on two matrix factorization problems, using a non-negative factorization model and a linear-Gaussian model." }, { "author": [ @@ -1788,14 +1788,14 @@ "Subramanian, Kaushik" ], "paper_title": "Apprenticeship Learning About Multiple Intentions", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Workshop on Nonlinear Speech Processing", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 897, + "page_form": 897, "page_to": 904, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper, we apply tools from inverse reinforcement learning (IRL) to the problem of learning from (unlabeled) demonstration trajectories of behavior generated by varying ``intentions'' or objectives. We derive an EM approach that clusters observed trajectories by inferring the objectives for each cluster using any of several possible IRL methods, and then uses the constructed clusters to quickly identify the intent of a trajectory. We show that a natural approach to IRL---a gradient ascent method that modifies reward parameters to maximize the likelihood of the observed trajectories---is successful at quickly identifying unknown reward functions. We demonstrate these ideas in the context of apprenticeship learning by acquiring the preferences of a human driver in a simple highway car simulator.\n" + "abstract": "Abstract:In this paper, we apply tools from inverse reinforcement learning (IRL) to the problem of learning from (unlabeled) demonstration trajectories of behavior generated by varying ``intentions'' or objectives. We derive an EM approach that clusters observed trajectories by inferring the objectives for each cluster using any of several possible IRL methods, and then uses the constructed clusters to quickly identify the intent of a trajectory. We show that a natural approach to IRL---a gradient ascent method that modifies reward parameters to maximize the likelihood of the observed trajectories---is successful at quickly identifying unknown reward functions. We demonstrate these ideas in the context of apprenticeship learning by acquiring the preferences of a human driver in a simple highway car simulator." }, { "author": [ @@ -1804,14 +1804,14 @@ "DeWeese, Michael" ], "paper_title": "Minimum Probability Flow Learning", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Conference of the Gesellschaft für Klassifikation", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 905, + "page_form": 905, "page_to": 912, "totle_page": 8, "language": "en", - "abstract": "Abstract:Fitting probabilistic models to data is often difficult, due to the general intractability of the partition function and its derivatives. Here we propose a new parameter estimation technique that does not require computing an intractable normalization factor or sampling from the equilibrium distribution of the model. This is achieved by establishing dynamics that would transfrom the observed data distribution into the model distribution, and then setting as the objective the minimization of the KL divergence between the data distribution and the distribution produced by running the dynamics for an infinitesimal time. Score matching, minimum velocity learning, and certain froms of contrastive divergence are shown to be special cases of this learning technique. We demonstrate parameter estimation in Ising models, deep belief networks and an independent component analysis model of natural scenes. In the Ising model case, current state of the art techniques are outperfromed by at least an order of magnitude in learning time, with lower error in recovered coupling parameters.\n" + "abstract": "Abstract:Fitting probabilistic models to data is often difficult, due to the general intractability of the partition function and its derivatives. Here we propose a new parameter estimation technique that does not require computing an intractable normalization factor or sampling from the equilibrium distribution of the model. This is achieved by establishing dynamics that would transform the observed data distribution into the model distribution, and then setting as the objective the minimization of the KL divergence between the data distribution and the distribution produced by running the dynamics for an infinitesimal time. Score matching, minimum velocity learning, and certain forms of contrastive divergence are shown to be special cases of this learning technique. We demonstrate parameter estimation in Ising models, deep belief networks and an independent component analysis model of natural scenes. In the Ising model case, current state of the art techniques are outperformed by at least an order of magnitude in learning time, with lower error in recovered coupling parameters." }, { "author": [ @@ -1821,14 +1821,14 @@ "Roy, Nicholas" ], "paper_title": "Infinite Dynamic Bayesian Networks", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Mobiles Computing in der Medizin", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 913, + "page_form": 913, "page_to": 920, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present the infinite dynamic Bayesian network model (iDBN), a nonparametric, factored state-space model that generalizes dynamic Bayesian networks (DBNs). The iDBN can infer every aspect of a DBN: the number of hidden factors, the number of values each factor can take, and (arbitrarily complex) connections and conditionals between factors and observations. In this way, the iDBN generalizes other nonparametric state space models, which until now generally focused on binary hidden nodes and more restricted connection structures. We show how this new prior allows us to find interesting structure in benchmark tests and on two real-world datasets involving weather data and neural infromation flow networks.\n" + "abstract": "Abstract:We present the infinite dynamic Bayesian network model (iDBN), a nonparametric, factored state-space model that generalizes dynamic Bayesian networks (DBNs). The iDBN can infer every aspect of a DBN: the number of hidden factors, the number of values each factor can take, and (arbitrarily complex) connections and conditionals between factors and observations. In this way, the iDBN generalizes other nonparametric state space models, which until now generally focused on binary hidden nodes and more restricted connection structures. We show how this new prior allows us to find interesting structure in benchmark tests and on two real-world datasets involving weather data and neural information flow networks." }, { "author": [ @@ -1836,28 +1836,28 @@ "Ng, Andrew" ], "paper_title": "The Importance of Encoding Versus Training with Sparse Coding and Vector Quantization ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Contemporary Computing in Ukraine", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 921, + "page_form": 921, "page_to": 928, "totle_page": 8, "language": "en", - "abstract": "Abstract: While vector quantization (VQ) has been applied widely to generate features for visual recognition problems, much recent work has focused on more powerful methods. In particular, sparse coding has emerged as a strong alternative to traditional VQ approaches and has been shown to achieve consistently higher perfromance on benchmark datasets. Both approaches can be split into a training phase, where the system learns a dictionary of basis functions from unlabeled data, and an encoding phase, where the dictionary is used to extract features from new inputs. In this work, we investigate the reasons for the success of sparse coding over VQ by decoupling these phases, allowing us to separate out the contributions of the training and encoding in a controlled way. Through extensive experiments on CIFAR, NORB and Caltech 101 datasets, we compare sparse coding and several other training and encoding schemes, including a from of VQ paired with a soft threshold activation function. Our results show not only that we can use fast VQ algorithms for training without penalty, but that we can just as well use randomly chosen exemplars from the training set. Rather than spend resources on training, we find it is more important to choose a good encoder---which can often be as simple as a feed forward non-linearity. Among our results, we demonstrate state-of-the-art perfromance on both CIFAR and NORB.\n" + "abstract": "Abstract: While vector quantization (VQ) has been applied widely to generate features for visual recognition problems, much recent work has focused on more powerful methods. In particular, sparse coding has emerged as a strong alternative to traditional VQ approaches and has been shown to achieve consistently higher performance on benchmark datasets. Both approaches can be split into a training phase, where the system learns a dictionary of basis functions from unlabeled data, and an encoding phase, where the dictionary is used to extract features from new inputs. In this work, we investigate the reasons for the success of sparse coding over VQ by decoupling these phases, allowing us to separate out the contributions of the training and encoding in a controlled way. Through extensive experiments on CIFAR, NORB and Caltech 101 datasets, we compare sparse coding and several other training and encoding schemes, including a form of VQ paired with a soft threshold activation function. Our results show not only that we can use fast VQ algorithms for training without penalty, but that we can just as well use randomly chosen exemplars from the training set. Rather than spend resources on training, we find it is more important to choose a good encoder---which can often be as simple as a feed forward non-linearity. Among our results, we demonstrate state-of-the-art performance on both CIFAR and NORB." }, { "author": [ "Cuturi, Marco" ], "paper_title": "Fast Global Alignment Kernels", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Datenverarbeitung im Marketing", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 929, + "page_form": 929, "page_to": 936, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose novel approaches to cast the widely-used family of Dynamic Time Warping (DTW) distances and similarities as positive definite kernels for time series. To this effect, we provide new theoretical insights on the family of Global Alignment kernels introduced by Cuturi et al. (2007) and propose alternative kernels which are both positive definite and faster to compute. We provide experimental evidence that these alternatives are both faster and more efficient in classification tasks than other kernels based on the DTW fromalism.\n" + "abstract": "Abstract:We propose novel approaches to cast the widely-used family of Dynamic Time Warping (DTW) distances and similarities as positive definite kernels for time series. To this effect, we provide new theoretical insights on the family of Global Alignment kernels introduced by Cuturi et al. (2007) and propose alternative kernels which are both positive definite and faster to compute. We provide experimental evidence that these alternatives are both faster and more efficient in classification tasks than other kernels based on the DTW formalism." }, { "author": [ @@ -1868,14 +1868,14 @@ "Ting, Jo-Anne" ], "paper_title": "Learning attentional policies for tracking and recognition in video with deep networks ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IEEE Transactions on Pattern Analysis and Machine Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 937, + "page_form": 937, "page_to": 944, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose a novel attentional model for simultaneous object tracking and recognition that is driven by gaze data. Motivated by theories of the human perceptual system, the model consists of two interacting pathways: ventral and dorsal. The ventral pathway models object appearance and classification using deep (factored)-restricted Boltzmann machines. At each point in time, the observations consist of retinal images, with decaying resolution toward the periphery of the gaze. The dorsal pathway models the location, orientation, scale and speed of the attended object. The posterior distribution of these states is estimated with particle filtering. Deeper in the dorsal pathway, we encounter an attentional mechanism that learns to control gazes so as to minimize tracking uncertainty. The approach is modular (with each module easily replaceable with more sophisticated algorithms), straightforward to implement, practically efficient, and works well in simple video sequences.\n" + "abstract": "Abstract:We propose a novel attentional model for simultaneous object tracking and recognition that is driven by gaze data. Motivated by theories of the human perceptual system, the model consists of two interacting pathways: ventral and dorsal. The ventral pathway models object appearance and classification using deep (factored)-restricted Boltzmann machines. At each point in time, the observations consist of retinal images, with decaying resolution toward the periphery of the gaze. The dorsal pathway models the location, orientation, scale and speed of the attended object. The posterior distribution of these states is estimated with particle filtering. Deeper in the dorsal pathway, we encounter an attentional mechanism that learns to control gazes so as to minimize tracking uncertainty. The approach is modular (with each module easily replaceable with more sophisticated algorithms), straightforward to implement, practically efficient, and works well in simple video sequences." }, { "author": [ @@ -1884,14 +1884,14 @@ "Bengio, Yoshua" ], "paper_title": "Large-Scale Learning of Embeddings with Reconstruction Sampling", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Neural Computation", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 945, + "page_form": 945, "page_to": 952, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper, we present a novel method to speed up the learning of embeddings for large-scale learning tasks involving very sparse data, as is typically the case for Natural Language Processing tasks. Our speed-up method has been developed in the context of Denoising Auto-encoders, which are trained in a purely unsupervised way to capture the input distribution, and learn embeddings for words and text similar to earlier neural language models. The main contribution is a new method to approximate reconstruction error by a sampling procedure. We show how this approximation can be made to obtain an unbiased estimator of the training criterion, and we show how it can be leveraged to make learning much more computationally efficient. We demonstrate the effectiveness of this method on the Amazon and RCV1 NLP datasets. Instead of reducing vocabulary size to make learning practical, our method allows us to train using very large vocabularies. In particular, reconstruction sampling requires 22x less training time on the full Amazon dataset.\n" + "abstract": "Abstract:In this paper, we present a novel method to speed up the learning of embeddings for large-scale learning tasks involving very sparse data, as is typically the case for Natural Language Processing tasks. Our speed-up method has been developed in the context of Denoising Auto-encoders, which are trained in a purely unsupervised way to capture the input distribution, and learn embeddings for words and text similar to earlier neural language models. The main contribution is a new method to approximate reconstruction error by a sampling procedure. We show how this approximation can be made to obtain an unbiased estimator of the training criterion, and we show how it can be leveraged to make learning much more computationally efficient. We demonstrate the effectiveness of this method on the Amazon and RCV1 NLP datasets. Instead of reducing vocabulary size to make learning practical, our method allows us to train using very large vocabularies. In particular, reconstruction sampling requires 22x less training time on the full Amazon dataset." }, { "author": [ @@ -1900,14 +1900,14 @@ "Chen, Yixin" ], "paper_title": "Automatic Feature Decomposition for Single View Co-training ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IEEE Transactions on Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 953, + "page_form": 953, "page_to": 960, "totle_page": 8, "language": "en", - "abstract": "Abstract:One of the most successful semi-supervised learning approaches is co-training for multi-view data. In co-training, one trains two classifiers, one for each view, and uses the most confident predictions of the unlabeled data for the two classifiers to ``teach each other''. In this paper, we extend co-training to learning scenarios without an explicit multi-view representation. Inspired by a theoretical analysis of Balcan et. al (2004), we introduce a novel algorithm that splits the feature space during learning, explicitly to encourage co-training to be successful. We demonstrate the efficacy of our proposed method in a weakly-supervised setting on the challenging Caltech-256 object recognition task, where we improve significantly over previous results by (Bergamo & Torresani, 2010) in almost all training-set size settings.\n" + "abstract": "Abstract:One of the most successful semi-supervised learning approaches is co-training for multi-view data. In co-training, one trains two classifiers, one for each view, and uses the most confident predictions of the unlabeled data for the two classifiers to ``teach each other''. In this paper, we extend co-training to learning scenarios without an explicit multi-view representation. Inspired by a theoretical analysis of Balcan et. al (2004), we introduce a novel algorithm that splits the feature space during learning, explicitly to encourage co-training to be successful. We demonstrate the efficacy of our proposed method in a weakly-supervised setting on the challenging Caltech-256 object recognition task, where we improve significantly over previous results by (Bergamo & Torresani, 2010) in almost all training-set size settings." }, { "author": [ @@ -1916,14 +1916,14 @@ "Kuboyama, Tetsuji" ], "paper_title": "Mapping kernels for trees ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Machine Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 961, + "page_form": 961, "page_to": 968, "totle_page": 8, "language": "en", - "abstract": "Abstract:We propose a comprehensive survey of tree kernels through the lens of the mapping kernels framework. We argue that most existing tree kernels, as well as many more that are presented for the first time in this paper, fall into a typology of kernels whose seemingly intricate computation can be efficiently factorized to yield polynomial time algorithms. Despite this fact, we argue that a naive implementation of such kernels remains prohibitively expensive to compute. We propose an approach whereby some computations for smaller trees are cached, which speeds up considerably the computation of all these tree kernels. We provide experimental evidence of this fact as well as preliminary results on the perfromance of these kernels.\n" + "abstract": "Abstract:We propose a comprehensive survey of tree kernels through the lens of the mapping kernels framework. We argue that most existing tree kernels, as well as many more that are presented for the first time in this paper, fall into a typology of kernels whose seemingly intricate computation can be efficiently factorized to yield polynomial time algorithms. Despite this fact, we argue that a naive implementation of such kernels remains prohibitively expensive to compute. We propose an approach whereby some computations for smaller trees are cached, which speeds up considerably the computation of all these tree kernels. We provide experimental evidence of this fact as well as preliminary results on the performance of these kernels." }, { "author": [ @@ -1934,14 +1934,14 @@ "Glotin, Herv\\'{e}" ], "paper_title": "Stochastic Low-Rank Kernel Learning for Regression ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 969, + "page_form": 969, "page_to": 976, "totle_page": 8, "language": "en", - "abstract": "Abstract:We present a novel approach to learn a kernel-based regression function. It is based on the use of conical combinations of data-based parameterized kernels and on a new stochastic convex optimization procedure of which we establish convergence guarantees. The overall learning procedure has the nice properties that a) the learned conical combination is automatically designed to perfrom the regression task at hand and b) the updates implicated by the optimization procedure are quite inexpensive. In order to shed light on the appositeness of our learning strategy, we present empirical results from experiments conducted on various benchmark datasets.\n" + "abstract": "Abstract:We present a novel approach to learn a kernel-based regression function. It is based on the use of conical combinations of data-based parameterized kernels and on a new stochastic convex optimization procedure of which we establish convergence guarantees. The overall learning procedure has the nice properties that a) the learned conical combination is automatically designed to perform the regression task at hand and b) the updates implicated by the optimization procedure are quite inexpensive. In order to shed light on the appositeness of our learning strategy, we present empirical results from experiments conducted on various benchmark datasets." }, { "author": [ @@ -1950,14 +1950,14 @@ "Aihara, Kazuyuki" ], "paper_title": "Size-constrained Submodular Minimization through Minimum Norm Base", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Journal of Machine Learning Research", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 977, + "page_form": 977, "page_to": 984, "totle_page": 8, "language": "en", - "abstract": "Abstract:A number of combinatorial optimization problems in machine learning can be described as the problem of minimizing a submodular function. It is known that the unconstrained submodular minimization problem can be solved in strongly polynomial time. However, additional constraints make the problem intractable in many settings. In this paper, we discuss the submodular minimization under a size constraint, which is NP-hard, and generalizes the densest subgraph problem and the unifrom graph partitioning problem. Because of NP-hardness, it is difficult to compute an optimal solution even for a prescribed size constraint. In our approach, we do not give approximation algorithms. Instead, the proposed algorithm computes optimal solutions for some of possible size constraints in polynomial time. Our algorithm utilizes the basic polyhedral theory associated with submodular functions. Additionally, we evaluate the perfromance of the proposed algorithm through computational experiments.\n" + "abstract": "Abstract:A number of combinatorial optimization problems in machine learning can be described as the problem of minimizing a submodular function. It is known that the unconstrained submodular minimization problem can be solved in strongly polynomial time. However, additional constraints make the problem intractable in many settings. In this paper, we discuss the submodular minimization under a size constraint, which is NP-hard, and generalizes the densest subgraph problem and the uniform graph partitioning problem. Because of NP-hardness, it is difficult to compute an optimal solution even for a prescribed size constraint. In our approach, we do not give approximation algorithms. Instead, the proposed algorithm computes optimal solutions for some of possible size constraints in polynomial time. Our algorithm utilizes the basic polyhedral theory associated with submodular functions. Additionally, we evaluate the performance of the proposed algorithm through computational experiments." }, { "author": [ @@ -1965,14 +1965,14 @@ "Torr, Philip" ], "paper_title": "Locally Linear Support Vector Machines ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Pattern Recognition Letters", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 985, + "page_form": 985, "page_to": 992, "totle_page": 8, "language": "en", - "abstract": "Abstract:Linear support vector machines ({\\sc svm}s) have become popular for solving classification tasks due to their fast and simple online application to large scale data sets. However, many problems are not linearly separable. For these problems kernel-based {\\sc svm}s are often used, but unlike their linear variant they suffer from various drawbacks in terms of computational and memory efficiency. Their response can be represented only as a function of the set of support vectors, which has been experimentally shown to grow linearly with the size of the training set. In this paper we propose a novel locally linear {\\sc svm} classifier with smooth decision boundary and bounded curvature. We show how the functions defining the classifier can be approximated using local codings and show how this model can be optimized in an online fashion by perfroming stochastic gradient descent with the same convergence guarantees as standard gradient descent method for linear {\\sc svm}. Our method achieves comparable perfromance to the state-of-the-art whilst being significantly faster than competing kernel {\\sc svm}s. We generalise this model to locally finite dimensional kernel {\\sc svm}.\n" + "abstract": "Abstract:Linear support vector machines ({\\sc svm}s) have become popular for solving classification tasks due to their fast and simple online application to large scale data sets. However, many problems are not linearly separable. For these problems kernel-based {\\sc svm}s are often used, but unlike their linear variant they suffer from various drawbacks in terms of computational and memory efficiency. Their response can be represented only as a function of the set of support vectors, which has been experimentally shown to grow linearly with the size of the training set. In this paper we propose a novel locally linear {\\sc svm} classifier with smooth decision boundary and bounded curvature. We show how the functions defining the classifier can be approximated using local codings and show how this model can be optimized in an online fashion by performing stochastic gradient descent with the same convergence guarantees as standard gradient descent method for linear {\\sc svm}. Our method achieves comparable performance to the state-of-the-art whilst being significantly faster than competing kernel {\\sc svm}s. We generalise this model to locally finite dimensional kernel {\\sc svm}." }, { "author": [ @@ -1983,14 +1983,14 @@ "Rakotomamonjy, Alain" ], "paper_title": "Functional Regularized Least Squares Classication with Operator-valued Kernels ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "The Journal of Artificial Societies and Social Simulation", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 993, + "page_form": 993, "page_to": 1000, "totle_page": 8, "language": "en", - "abstract": "Abstract:Although operator-valued kernels have recently received increasing interest in various machine learning and functional data analysis problems such as multi-task learning or functional regression, little attention has been paid to the understanding of their associated feature spaces. In this paper, we explore the potential of adopting an operator-valued kernel feature space perspective for the analysis of functional data. We then extend the Regularized Least Squares Classification (RLSC) algorithm to cover situations where there are multiple functions per observation. Experiments on a sound recognition problem show that the proposed method outperfroms the classical RLSC algorithm.\n" + "abstract": "Abstract:Although operator-valued kernels have recently received increasing interest in various machine learning and functional data analysis problems such as multi-task learning or functional regression, little attention has been paid to the understanding of their associated feature spaces. In this paper, we explore the potential of adopting an operator-valued kernel feature space perspective for the analysis of functional data. We then extend the Regularized Least Squares Classification (RLSC) algorithm to cover situations where there are multiple functions per observation. Experiments on a sound recognition problem show that the proposed method outperforms the classical RLSC algorithm." }, { "author": [ @@ -2000,14 +2000,14 @@ "Xu, Huan" ], "paper_title": "Clustering Partially Observed Graphs via Convex Optimization ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal of Pattern Recognition and Artificial Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1001, + "page_form": 1001, "page_to": 1008, "totle_page": 8, "language": "en", - "abstract": "Abstract:This paper considers the problem of clustering a partially observed unweighted graph -- i.e. one where for some node pairs we know there is an edge between them, for some others we know there is no edge, and for the remaining we do not know whether or not there is an edge. We want to organize the nodes into disjoint clusters so that there is relatively dense (observed) connectivity within clusters, and sparse across clusters. We take a novel yet natural approach to this problem, by focusing on finding the clustering that minimizes the number of \"disagreements\" - i.e. the sum of the number of (observed) missing edges within clusters, and (observed) present edges across clusters. Our algorithm uses convex optimization; its basis is a reduction of disagreement minimization to the problem of recovering an (unknown) low-rank matrix and an (unknown) sparse matrix from their partially observed sum. We show that our algorithm succeeds under certain natural assumptions on the optimal clustering and its disagreements. Our results significantly strengthen existing matrix splitting results for the special case of our clustering problem. Our results directly enhance solutions to the problem of Correlation Clustering with partial observations\n" + "abstract": "Abstract:This paper considers the problem of clustering a partially observed unweighted graph -- i.e. one where for some node pairs we know there is an edge between them, for some others we know there is no edge, and for the remaining we do not know whether or not there is an edge. We want to organize the nodes into disjoint clusters so that there is relatively dense (observed) connectivity within clusters, and sparse across clusters. We take a novel yet natural approach to this problem, by focusing on finding the clustering that minimizes the number of \"disagreements\" - i.e. the sum of the number of (observed) missing edges within clusters, and (observed) present edges across clusters. Our algorithm uses convex optimization; its basis is a reduction of disagreement minimization to the problem of recovering an (unknown) low-rank matrix and an (unknown) sparse matrix from their partially observed sum. We show that our algorithm succeeds under certain natural assumptions on the optimal clustering and its disagreements. Our results significantly strengthen existing matrix splitting results for the special case of our clustering problem. Our results directly enhance solutions to the problem of Correlation Clustering with partial observations" }, { "author": [ @@ -2015,14 +2015,14 @@ "Ravikumar, Pradeep" ], "paper_title": "On the Use of Variational Inference for Learning Discrete Graphical Model", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Information Research", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1009, + "page_form": 1009, "page_to": 1016, "totle_page": 8, "language": "en", - "abstract": "Abstract:We study the general class of estimators for graphical model structure based on optimizing $\\ell_1$-regularized approximate log-likelihood, where the approximate likelihood uses tractable variational approximations of the partition function. We provide a message-passing algorithm that \\emph{directly} computes the $\\ell_1$ regularized approximate MLE. Further, in the case of certain reweighted entropy approximations to the partition function, we show that surprisingly the $\\ell_1$ regularized approximate MLE estimator has a \\emph{closed-from}, so that we would no longer need to run through many iterations of approximate inference and message-passing. Lastly, we analyze this general class of estimators for graph structure recovery, or its \\emph{sparsistency}, and show that it is indeed sparsistent under certain conditions.\n" + "abstract": "Abstract:We study the general class of estimators for graphical model structure based on optimizing $\\ell_1$-regularized approximate log-likelihood, where the approximate likelihood uses tractable variational approximations of the partition function. We provide a message-passing algorithm that \\emph{directly} computes the $\\ell_1$ regularized approximate MLE. Further, in the case of certain reweighted entropy approximations to the partition function, we show that surprisingly the $\\ell_1$ regularized approximate MLE estimator has a \\emph{closed-form}, so that we would no longer need to run through many iterations of approximate inference and message-passing. Lastly, we analyze this general class of estimators for graph structure recovery, or its \\emph{sparsistency}, and show that it is indeed sparsistent under certain conditions." }, { "author": [ @@ -2031,14 +2031,14 @@ "Hinton, Geoffrey" ], "paper_title": "Generating Text with Recurrent Neural Networks ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Advanced Robotics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1017, + "page_form": 1017, "page_to": 1024, "totle_page": 8, "language": "en", - "abstract": "Abstract:Recurrent Neural Networks (RNNs) are very powerful sequence models that do not enjoy widespread use because it is extremely difficult to train them properly. Fortunately, recent advances in Hessian-free optimization have been able to overcome the difficulties associated with training RNNs, making it possible to apply them successfully to challenging sequence problems. In this paper we demonstrate the power of RNNs trained with the new Hessian-Free optimizer (HF) by applying them to character-level language modeling tasks. The standard RNN architecture, while effective, is not ideally suited for such tasks, so we introduce a new RNN variant that uses multiplicative (or ``gated'') connections which allow the current input character to determine the transition matrix from one hidden state vector to the next. After training the multiplicative RNN with the HF optimizer for five days on 8 high-end Graphics Processing Units, we were able to surpass the perfromance of the best previous single method for character-level language modeling -- a hierarchical non-parametric sequence model. To our knowledge this represents the largest recurrent neural network application to date.\n" + "abstract": "Abstract:Recurrent Neural Networks (RNNs) are very powerful sequence models that do not enjoy widespread use because it is extremely difficult to train them properly. Fortunately, recent advances in Hessian-free optimization have been able to overcome the difficulties associated with training RNNs, making it possible to apply them successfully to challenging sequence problems. In this paper we demonstrate the power of RNNs trained with the new Hessian-Free optimizer (HF) by applying them to character-level language modeling tasks. The standard RNN architecture, while effective, is not ideally suited for such tasks, so we introduce a new RNN variant that uses multiplicative (or ``gated'') connections which allow the current input character to determine the transition matrix from one hidden state vector to the next. After training the multiplicative RNN with the HF optimizer for five days on 8 high-end Graphics Processing Units, we were able to surpass the performance of the best previous single method for character-level language modeling -- a hierarchical non-parametric sequence model. To our knowledge this represents the largest recurrent neural network application to date." }, { "author": [ @@ -2047,14 +2047,14 @@ "Chatterje, Snigdhansu" ], "paper_title": "Probabilistic Matrix Addition", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Information Systems Frontiers", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1025, + "page_form": 1025, "page_to": 1032, "totle_page": 8, "language": "en", - "abstract": "Abstract:We introduce Probabilistic Matrix Addition (PMA) for modeling real-valued data matrices by simultaneously capturing covariance structure among rows and among columns. PMA additively combines two latent matrices drawn from two Gaussian Processes respectively over rows and columns. The resulting joint distribution over the observed matrix does not factorize over entries, rows, or columns, and can thus capture intricate dependencies in the matrix. Exact inference in PMA is possible, but involves inversion of large matrices, and can be computationally prohibitive. Efficient approximate inference is possible due to the sparse dependency structure among latent variables. We propose two families of approximate inference algorithms for PMA based on Gibbs sampling and MAP inference. We demonstrate the effectiveness of PMA for missing value prediction and multi-label classification problems.\n" + "abstract": "Abstract:We introduce Probabilistic Matrix Addition (PMA) for modeling real-valued data matrices by simultaneously capturing covariance structure among rows and among columns. PMA additively combines two latent matrices drawn from two Gaussian Processes respectively over rows and columns. The resulting joint distribution over the observed matrix does not factorize over entries, rows, or columns, and can thus capture intricate dependencies in the matrix. Exact inference in PMA is possible, but involves inversion of large matrices, and can be computationally prohibitive. Efficient approximate inference is possible due to the sparse dependency structure among latent variables. We propose two families of approximate inference algorithms for PMA based on Gibbs sampling and MAP inference. We demonstrate the effectiveness of PMA for missing value prediction and multi-label classification problems." }, { "author": [ @@ -2062,14 +2062,14 @@ "Sutskever, Ilya" ], "paper_title": "Learning Recurrent Neural Networks with Hessian-Free Optimization ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Pattern Analysis and Applications", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1033, + "page_form": 1033, "page_to": 1040, "totle_page": 8, "language": "en", - "abstract": "Abstract: In this work we resolve the long-outstanding problem of how to effectively train recurrent neural networks (RNNs) on complex and difficult sequence modeling problems which may contain long-term data dependencies. Utilizing recent advances in the Hessian-free optimization approach \\citep{hf}, together with a novel damping scheme, we successfully train RNNs on two sets of challenging problems. First, a collection of pathological synthetic datasets which are known to be impossible for standard optimization approaches (due to their extremely long-term dependencies), and second, on three natural and highly complex real-world sequence datasets where we find that our method significantly outperfroms the previous state-of-the-art method for training neural sequence models: the Long Short-term Memory approach of \\citet{lstm}. Additionally, we offer a new interpretation of the generalized Gauss-Newton matrix of \\citet{schraudolph} which is used within the HF approach of Martens.\n" + "abstract": "Abstract: In this work we resolve the long-outstanding problem of how to effectively train recurrent neural networks (RNNs) on complex and difficult sequence modeling problems which may contain long-term data dependencies. Utilizing recent advances in the Hessian-free optimization approach \\citep{hf}, together with a novel damping scheme, we successfully train RNNs on two sets of challenging problems. First, a collection of pathological synthetic datasets which are known to be impossible for standard optimization approaches (due to their extremely long-term dependencies), and second, on three natural and highly complex real-world sequence datasets where we find that our method significantly outperforms the previous state-of-the-art method for training neural sequence models: the Long Short-term Memory approach of \\citet{lstm}. Additionally, we offer a new interpretation of the generalized Gauss-Newton matrix of \\citet{schraudolph} which is used within the HF approach of Martens." }, { "author": [ @@ -2078,14 +2078,14 @@ "Xing, Eric" ], "paper_title": "Sparse Additive Generative Models of Text ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal on Document Analysis and Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1041, + "page_form": 1041, "page_to": 1048, "totle_page": 8, "language": "en", - "abstract": "Abstract:Generative models of text typically associate a multinomial with every class label or topic. Even in simple models this requires the estimation of thousands of parameters; in multifaceted latent variable models, standard approaches require additional latent ``switching'' variables for every token, complicating inference. In this paper, we propose an alternative generative model for text. The central idea is that each class label or latent topic is endowed with a model of the deviation in log-frequency from a constant background distribution. This approach has two key advantages: we can enforce sparsity to prevent overfitting, and we can combine generative facets through simple addition in log space, avoiding the need for latent switching variables. We demonstrate the applicability of this idea to a range of scenarios: classification, topic modeling, and more complex multifaceted generative models.\n" + "abstract": "Abstract:Generative models of text typically associate a multinomial with every class label or topic. Even in simple models this requires the estimation of thousands of parameters; in multifaceted latent variable models, standard approaches require additional latent ``switching'' variables for every token, complicating inference. In this paper, we propose an alternative generative model for text. The central idea is that each class label or latent topic is endowed with a model of the deviation in log-frequency from a constant background distribution. This approach has two key advantages: we can enforce sparsity to prevent overfitting, and we can combine generative facets through simple addition in log space, avoiding the need for latent switching variables. We demonstrate the applicability of this idea to a range of scenarios: classification, topic modeling, and more complex multifaceted generative models." }, { "author": [ @@ -2095,14 +2095,14 @@ "Scherrer, Bruno" ], "paper_title": "Classification-based Policy Iteration with a Critic", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Neural Computing and Applications", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1049, + "page_form": 1049, "page_to": 1056, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper, we study the effect of adding a value function approximation component (critic) to rollout classification-based policy iteration (RCPI) algorithms. The idea is to use a critic to approximate the return after we truncate the rollout trajectories. This allows us to control the bias and variance of the rollout estimates of the action-value function. Therefore, the introduction of a critic can improve the accuracy of the rollout estimates, and as a result, enhance the perfromance of the RCPI algorithm. We present a new RCPI algorithm, called direct policy iteration with critic (DPI-Critic), and provide its finite-sample analysis when the critic is based on the LSTD method. We empirically evaluate the perfromance of DPI-Critic and compare it with DPI and LSPI in two benchmark reinforcement learning problems.\n" + "abstract": "Abstract:In this paper, we study the effect of adding a value function approximation component (critic) to rollout classification-based policy iteration (RCPI) algorithms. The idea is to use a critic to approximate the return after we truncate the rollout trajectories. This allows us to control the bias and variance of the rollout estimates of the action-value function. Therefore, the introduction of a critic can improve the accuracy of the rollout estimates, and as a result, enhance the performance of the RCPI algorithm. We present a new RCPI algorithm, called direct policy iteration with critic (DPI-Critic), and provide its finite-sample analysis when the critic is based on the LSTD method. We empirically evaluate the performance of DPI-Critic and compare it with DPI and LSPI in two benchmark reinforcement learning problems." }, { "author": [ @@ -2110,14 +2110,14 @@ "Kempe, David" ], "paper_title": "Submodular meets Spectral: Greedy Algorithms for Subset Selection, Sparse Approximation and Dictionary Selection ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "European Journal of Information Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1057, + "page_form": 1057, "page_to": 1064, "totle_page": 8, "language": "en", - "abstract": "Abstract:We study the problem of selecting a subset of $k$ random variables from a large set, in order to obtain the best linear prediction of another variable of interest. This problem can be viewed in the context of both feature selection and sparse approximation. We analyze the perfromance of widely used greedy heuristics, using insights from the maximization of submodular functions and spectral analysis. We introduce the submodularity ratio as a key quantity to help understand why greedy algorithms perfrom well even when the variables are highly correlated. Using our techniques, we obtain the strongest known approximation guarantees for this problem, both in terms of the submodularity ratio and the smallest $k$-sparse eigenvalue of the covariance matrix. We also analyze greedy algorithms for the dictionary selection problem, and significantly improve the previously known guarantees. Our theoretical analysis is complemented by experiments on real-world and synthetic data sets; the experiments show that the submodular ratio is a stronger predictor of the perfromance of greedy algorithms than other spectral parameters.\n" + "abstract": "Abstract:We study the problem of selecting a subset of $k$ random variables from a large set, in order to obtain the best linear prediction of another variable of interest. This problem can be viewed in the context of both feature selection and sparse approximation. We analyze the performance of widely used greedy heuristics, using insights from the maximization of submodular functions and spectral analysis. We introduce the submodularity ratio as a key quantity to help understand why greedy algorithms perform well even when the variables are highly correlated. Using our techniques, we obtain the strongest known approximation guarantees for this problem, both in terms of the submodularity ratio and the smallest $k$-sparse eigenvalue of the covariance matrix. We also analyze greedy algorithms for the dictionary selection problem, and significantly improve the previously known guarantees. Our theoretical analysis is complemented by experiments on real-world and synthetic data sets; the experiments show that the submodular ratio is a stronger predictor of the performance of greedy algorithms than other spectral parameters." }, { "author": [ @@ -2126,14 +2126,14 @@ "Xing, Eric" ], "paper_title": "A Spectral Algorithm for Latent Tree Graphical Models", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Internet Mathematics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1065, + "page_form": 1065, "page_to": 1072, "totle_page": 8, "language": "en", - "abstract": "Abstract:Latent variable models are powerful tools for probabilistic modeling, and have been successfully applied to various domains, such as speech analysis and bioinfromatics. However, parameter learning algorithms for latent variable models have predominantly relied on local search heuristics such as expectation maximization (EM). We propose a fast, local-minimum-free spectral algorithm for learning latent variable models with arbitrary tree topologies, and show that the joint distribution of the observed variables can be reconstructed from the marginals of triples of observed variables irrespective of the maximum degree of the tree. We demonstrate the perfromance of our spectral algorithm on synthetic and real datasets; for large training sizes, our algorithm perfroms comparable to or better than EM while being orders of magnitude faster.\n" + "abstract": "Abstract:Latent variable models are powerful tools for probabilistic modeling, and have been successfully applied to various domains, such as speech analysis and bioinformatics. However, parameter learning algorithms for latent variable models have predominantly relied on local search heuristics such as expectation maximization (EM). We propose a fast, local-minimum-free spectral algorithm for learning latent variable models with arbitrary tree topologies, and show that the joint distribution of the observed variables can be reconstructed from the marginals of triples of observed variables irrespective of the maximum degree of the tree. We demonstrate the performance of our spectral algorithm on synthetic and real datasets; for large training sizes, our algorithm performs comparable to or better than EM while being orders of magnitude faster." }, { "author": [ @@ -2142,14 +2142,14 @@ "Jordan, Michael" ], "paper_title": "A Unified Probabilistic Model for Global and Local Unsupervised Feature Selection", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Transactions on Rough Sets", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1073, + "page_form": 1073, "page_to": 1080, "totle_page": 8, "language": "en", - "abstract": "Abstract:Existing algorithms for joint clustering and feature selection can be categorized as either global or local approaches. Global methods select a single cluster-independent subset of features, whereas local methods select cluster-specific subsets of features. In this paper, we present a unified probabilistic model that can perfrom both global and local feature selection for clustering. Our approach is based on a hierarchical beta-Bernoulli prior combined with a Dirichlet process mixture model. We obtain global or local feature selection by adjusting the variance of the beta prior. We provide a variational inference algorithm for our model. In addition to simultaneously learning the clusters and features, this Bayesian fromulation allows us to learn both the number of clusters and the number of features to retain. Experiments on synthetic and real data show that our unified model can find global and local features and cluster data as well as competing methods of each type.\n" + "abstract": "Abstract:Existing algorithms for joint clustering and feature selection can be categorized as either global or local approaches. Global methods select a single cluster-independent subset of features, whereas local methods select cluster-specific subsets of features. In this paper, we present a unified probabilistic model that can perform both global and local feature selection for clustering. Our approach is based on a hierarchical beta-Bernoulli prior combined with a Dirichlet process mixture model. We obtain global or local feature selection by adjusting the variance of the beta prior. We provide a variational inference algorithm for our model. In addition to simultaneously learning the clusters and features, this Bayesian formulation allows us to learn both the number of clusters and the number of features to retain. Experiments on synthetic and real data show that our unified model can find global and local features and cluster data as well as competing methods of each type." }, { "author": [ @@ -2157,14 +2157,14 @@ "Zhou, Zhi-Hua" ], "paper_title": "Towards Making Unlabeled Data Never Hurt ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IMPACT Impact of Computing in Science and Engineering", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1081, + "page_form": 1081, "page_to": 1088, "totle_page": 8, "language": "en", - "abstract": "Abstract:It is usually expected that, when labeled data are limited, the learning perfromance can be improved by exploiting unlabeled data. In many cases, however, the perfromances of current semi-supervised learning approaches may be even worse than purely using the limited labeled data.It is desired to have \\textit{safe} semi-supervised learning approaches which never degenerate learning perfromance by using unlabeled data. In this paper, we focus on semi-supervised support vector machines (S3VMs) and propose S4VMs, i.e., safe S3VMs. Unlike S3VMs which typically aim at approaching an optimal low-density separator, S4VMs try to exploit the candidate low-density separators simultaneously to reduce the risk of identifying a poor separator with unlabeled data. We describe two implementations of S4VMs, and our comprehensive experiments show that the overall perfromance of S4VMs are highly competitive to S3VMs, while in contrast to S3VMs which degenerate perfromance in many cases, S4VMs are never significantly inferior to inductive SVMs.\n" + "abstract": "Abstract:It is usually expected that, when labeled data are limited, the learning performance can be improved by exploiting unlabeled data. In many cases, however, the performances of current semi-supervised learning approaches may be even worse than purely using the limited labeled data.It is desired to have \\textit{safe} semi-supervised learning approaches which never degenerate learning performance by using unlabeled data. In this paper, we focus on semi-supervised support vector machines (S3VMs) and propose S4VMs, i.e., safe S3VMs. Unlike S3VMs which typically aim at approaching an optimal low-density separator, S4VMs try to exploit the candidate low-density separators simultaneously to reduce the risk of identifying a poor separator with unlabeled data. We describe two implementations of S4VMs, and our comprehensive experiments show that the overall performance of S4VMs are highly competitive to S3VMs, while in contrast to S3VMs which degenerate performance in many cases, S4VMs are never significantly inferior to inductive SVMs." }, { "author": [ @@ -2176,14 +2176,14 @@ "Ng, Andrew" ], "paper_title": "On Random Weights and Unsupervised Feature Learning ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Cognition, Technology & Work", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1089, + "page_form": 1089, "page_to": 1096, "totle_page": 8, "language": "en", - "abstract": "Abstract:Recently two anomalous results in the literature have shown that certain feature learning architectures can yield useful features for object recognition tasks even with untrained, random weights. In this paper we pose the question: why do random weights sometimes do so well? Our answer is that certain convolutional pooling architectures can be inherently frequency selective and translation invariant, even with random weights. Based on this we demonstrate the viability of extremely fast architecture search by u\n" + "abstract": "Abstract:Recently two anomalous results in the literature have shown that certain feature learning architectures can yield useful features for object recognition tasks even with untrained, random weights. In this paper we pose the question: why do random weights sometimes do so well? Our answer is that certain convolutional pooling architectures can be inherently frequency selective and translation invariant, even with random weights. Based on this we demonstrate the viability of extremely fast architecture search by u" }, { "author": [ @@ -2192,14 +2192,14 @@ "Li, Lihong" ], "paper_title": "Doubly Robust Policy Evaluation and Learning ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal of Information Technology and Decision Making", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1097, + "page_form": 1097, "page_to": 1104, "totle_page": 8, "language": "en", - "abstract": "Abstract: We study decision making in environments where the reward is only partially observed, but can be modeled as a function of an action and an observed context. This setting, known as contextual bandits, encompasses a wide variety of applications including health-care policy and Internet advertising. A central task is evaluation of a new policy given historic data consisting of contexts, actions and received rewards. The key challenge is that the past data typically does not faithfully represent proportions of actions taken by a new policy. Previous approaches rely either on models of rewards or models of the past policy. The fromer are plagued by a large bias whereas the latter have a large variance. In this work, we leverage the strength and overcome the weaknesses of the two approaches by applying the \\emph{doubly robust} technique to the problems of policy evaluation and optimization. We prove that this approach yields accurate value estimates when we have \\emph{either} a good (but not necessarily consistent) model of rewards \\emph{or} a good (but not necessarily consistent) model of past policy. Extensive empirical comparison demonstrates that the doubly robust approach unifromly improves over existing techniques, achieving both lower variance in value estimation and better policies. As such, we expect the doubly robust approach to become common practice.\n" + "abstract": "Abstract: We study decision making in environments where the reward is only partially observed, but can be modeled as a function of an action and an observed context. This setting, known as contextual bandits, encompasses a wide variety of applications including health-care policy and Internet advertising. A central task is evaluation of a new policy given historic data consisting of contexts, actions and received rewards. The key challenge is that the past data typically does not faithfully represent proportions of actions taken by a new policy. Previous approaches rely either on models of rewards or models of the past policy. The former are plagued by a large bias whereas the latter have a large variance. In this work, we leverage the strength and overcome the weaknesses of the two approaches by applying the \\emph{doubly robust} technique to the problems of policy evaluation and optimization. We prove that this approach yields accurate value estimates when we have \\emph{either} a good (but not necessarily consistent) model of rewards \\emph{or} a good (but not necessarily consistent) model of past policy. Extensive empirical comparison demonstrates that the doubly robust approach uniformly improves over existing techniques, achieving both lower variance in value estimation and better policies. As such, we expect the doubly robust approach to become common practice." }, { "author": [ @@ -2209,14 +2209,14 @@ "Ng, Andrew" ], "paper_title": "Learning Deep Energy Models", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal of Knowledge and Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1105, + "page_form": 1105, "page_to": 1112, "totle_page": 8, "language": "en", - "abstract": "Abstract:Deep generative models with multiple hidden layers have been shown to be able to learn meaningful and compact representations of data. In this work we propose deep energy models, a class of models that use a deep feedforward neural network to model the energy landscape that defines a probabilistic model. We are able to efficiently train all layers of our model at the same time, allowing the lower layers of the model to adapt to the training of the higher layers, producing better generative models. We evaluate the generative perfromance of our models on natural images and demonstrate that joint training of multiple layers yields qualitative and quantitative improvements over greedy layerwise training. We further generalize our models beyond the commonly used sigmoidal neural networks and show how a deep extension of the product of Student-t distributions model achieves good generative perfromance. Finally, we introduce a discriminative extension of our model and demonstrate that it outperfroms other fully-connected models on object recognition on the NORB dataset.\n" + "abstract": "Abstract:Deep generative models with multiple hidden layers have been shown to be able to learn meaningful and compact representations of data. In this work we propose deep energy models, a class of models that use a deep feedforward neural network to model the energy landscape that defines a probabilistic model. We are able to efficiently train all layers of our model at the same time, allowing the lower layers of the model to adapt to the training of the higher layers, producing better generative models. We evaluate the generative performance of our models on natural images and demonstrate that joint training of multiple layers yields qualitative and quantitative improvements over greedy layerwise training. We further generalize our models beyond the commonly used sigmoidal neural networks and show how a deep extension of the product of Student-t distributions model achieves good generative performance. Finally, we introduce a discriminative extension of our model and demonstrate that it outperforms other fully-connected models on object recognition on the NORB dataset." }, { "author": [ @@ -2225,14 +2225,14 @@ "Huellermeier, Eyke" ], "paper_title": "Bipartite Ranking through Minimization of Univariate Loss", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Journal of Web and Grid Services", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1113, + "page_form": 1113, "page_to": 1120, "totle_page": 8, "language": "en", - "abstract": "Abstract:Minimization of the rank loss or, equivalently, maximization of the AUC in bipartite ranking calls for minimizing the number of disagreements between pairs of instances. Since the complexity of this problem is inherently quadratic in the number of training examples, it is tempting to ask how much is actually lost by minimizing a simple univariate loss function, as done by standard classification methods, as a surrogate. In this paper, we first note that minimization of 0/1 loss is not an option, as it may yield an arbitrarily high rank loss. We show, however, that better results can be achieved by means of a weighted (cost-sensitive) version of 0/1 loss. Yet, the real gain is obtained through margin-based loss functions, for which we are able to derive proper bounds, not only for rank risk but, more importantly, also for rank regret. The paper is completed with an experimental study in which we address specific questions raised by our theoretical analysis.\n" + "abstract": "Abstract:Minimization of the rank loss or, equivalently, maximization of the AUC in bipartite ranking calls for minimizing the number of disagreements between pairs of instances. Since the complexity of this problem is inherently quadratic in the number of training examples, it is tempting to ask how much is actually lost by minimizing a simple univariate loss function, as done by standard classification methods, as a surrogate. In this paper, we first note that minimization of 0/1 loss is not an option, as it may yield an arbitrarily high rank loss. We show, however, that better results can be achieved by means of a weighted (cost-sensitive) version of 0/1 loss. Yet, the real gain is obtained through margin-based loss functions, for which we are able to derive proper bounds, not only for rank risk but, more importantly, also for rank regret. The paper is completed with an experimental study in which we address specific questions raised by our theoretical analysis." }, { "author": [ @@ -2240,14 +2240,14 @@ "Wright, Stephen" ], "paper_title": "Manifold Identification of Dual Averaging Methods for Regularized Stochastic Online Learning", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Robotica", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1121, + "page_form": 1121, "page_to": 1128, "totle_page": 8, "language": "en", - "abstract": "Abstract:Iterative methods that take steps in approximate subgradient directions have proved to be useful for stochastic learning problems over large or streaming data sets. When the objective consists of a loss function plus a nonsmooth regularization term, whose purpose is to induce structure (for example, sparsity) in the solution, the solution often lies on a low-dimensional manifold along which the regularizer is smooth. This paper shows that a regularized dual averaging algorithm can identify this manifold with high probability. This observation motivates an algorithmic strategy in which, once a near-optimal manifold is identified, we switch to an algorithm that searches only in this manifold, which typically has much lower intrinsic dimension than the full space, thus converging quickly to a near-optimal point with the desired structure. Computational results are presented to illustrate these claims.\n" + "abstract": "Abstract:Iterative methods that take steps in approximate subgradient directions have proved to be useful for stochastic learning problems over large or streaming data sets. When the objective consists of a loss function plus a nonsmooth regularization term, whose purpose is to induce structure (for example, sparsity) in the solution, the solution often lies on a low-dimensional manifold along which the regularizer is smooth. This paper shows that a regularized dual averaging algorithm can identify this manifold with high probability. This observation motivates an algorithmic strategy in which, once a near-optimal manifold is identified, we switch to an algorithm that searches only in this manifold, which typically has much lower intrinsic dimension than the full space, thus converging quickly to a near-optimal point with the desired structure. Computational results are presented to illustrate these claims." }, { "author": [ @@ -2256,14 +2256,14 @@ "Wainwright, Martin" ], "paper_title": "Noisy matrix decomposition via convex relaxation: Optimal rates in high dimensions ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Machine Intelligence and Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1129, + "page_form": 1129, "page_to": 1136, "totle_page": 8, "language": "en", - "abstract": "Abstract:We analyze a class of estimators based on a convex relaxation for solving high-dimensional matrix decomposition problems. The observations are the noisy realizations of the sum of an (approximately) low rank matrix $\\Theta^\\star$ with a second matrix $\\Gamma^\\star$ endowed with a complementary from of low-dimensional structure. We derive a general theorem that gives upper bounds on the Frobenius norm error for an estimate of the pair $(\\Theta^\\star, \\Gamma^\\star)$ obtained by solving a convex optimization problem. We then specialize our general result to two cases that have been studied in the context of robust PCA: low rank plus sparse structure, and low rank plus a column sparse structure. Our theory yields Frobenius norm error bounds for both deterministic and stochastic noise matrices, and in the latter case, they are minimax optimal. The sharpness of our theoretical predictions is also confirmed in numerical simulations.\n" + "abstract": "Abstract:We analyze a class of estimators based on a convex relaxation for solving high-dimensional matrix decomposition problems. The observations are the noisy realizations of the sum of an (approximately) low rank matrix $\\Theta^\\star$ with a second matrix $\\Gamma^\\star$ endowed with a complementary form of low-dimensional structure. We derive a general theorem that gives upper bounds on the Frobenius norm error for an estimate of the pair $(\\Theta^\\star, \\Gamma^\\star)$ obtained by solving a convex optimization problem. We then specialize our general result to two cases that have been studied in the context of robust PCA: low rank plus sparse structure, and low rank plus a column sparse structure. Our theory yields Frobenius norm error bounds for both deterministic and stochastic noise matrices, and in the latter case, they are minimax optimal. The sharpness of our theoretical predictions is also confirmed in numerical simulations." }, { "author": [ @@ -2272,14 +2272,14 @@ "Mannor, Shie" ], "paper_title": "Bundle Selling by Online Estimation of Valuation Functions", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "IEEE Transactions on Aerospace and Navigational Electronics", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1137, + "page_form": 1137, "page_to": 1144, "totle_page": 8, "language": "en", - "abstract": "Abstract:We consider the problem of online selection of a bundle of items when the cost of each item changes arbitrarily from round to round and the valuation function is initially unknown and revealed only through the noisy values of selected bundles (the bandit feedback setting). We are interested in learning schemes that have a small regret compared to an agent who knows the true valuation function. Since there are exponentially many bundles, further assumptions on the valuation functions are needed. We make the assumption that the valuation function is supermodular and has non-linear interactions that are of low degree in a novel sense. We develop efficient learning algorithms that balance exploration and exploitation to achieve low regret in this setting.\n" + "abstract": "Abstract:We consider the problem of online selection of a bundle of items when the cost of each item changes arbitrarily from round to round and the valuation function is initially unknown and revealed only through the noisy values of selected bundles (the bandit feedback setting). We are interested in learning schemes that have a small regret compared to an agent who knows the true valuation function. Since there are exponentially many bundles, further assumptions on the valuation functions are needed. We make the assumption that the valuation function is supermodular and has non-linear interactions that are of low degree in a novel sense. We develop efficient learning algorithms that balance exploration and exploitation to achieve low regret in this setting." }, { "author": [ @@ -2288,14 +2288,14 @@ "Bengio, Yoshua" ], "paper_title": "Unsupervised Models of Images by Spike-and-Slab RBMs", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Transactions on Machine Learning and Data Mining", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1145, + "page_form": 1145, "page_to": 1152, "totle_page": 8, "language": "en", - "abstract": "Abstract:The spike and slab Restricted Boltzmann Machine (RBM) is defined by having both a real valued ``slab'' variable and a binary ``spike'' variable associated with each unit in the hidden layer. In this paper we generalize and extend the spike and slab RBM to include non-zero means of the conditional distribution over the observed variables conditional on the binary spike variables. We also introduce a term, quadratic in the observed data that we exploit to guarantee the all conditionals associated with the model are well defined -- a guarantee that was absent in the original spike and slab RBM. The inclusion of these generalizations improves the perfromance of the spike and slab RBM as a feature learner and achieves competitive perfromance on the CIFAR-10 image classification task. The spike and slab model, when trained in a convolutional configuration, can generate sensible samples that demonstrate that the model has capture the broad statistical structure of natural images.\n" + "abstract": "Abstract:The spike and slab Restricted Boltzmann Machine (RBM) is defined by having both a real valued ``slab'' variable and a binary ``spike'' variable associated with each unit in the hidden layer. In this paper we generalize and extend the spike and slab RBM to include non-zero means of the conditional distribution over the observed variables conditional on the binary spike variables. We also introduce a term, quadratic in the observed data that we exploit to guarantee the all conditionals associated with the model are well defined -- a guarantee that was absent in the original spike and slab RBM. The inclusion of these generalizations improves the performance of the spike and slab RBM as a feature learner and achieves competitive performance on the CIFAR-10 image classification task. The spike and slab model, when trained in a convolutional configuration, can generate sensible samples that demonstrate that the model has capture the broad statistical structure of natural images." }, { "author": [ @@ -2304,14 +2304,14 @@ "Langmead, Christopher" ], "paper_title": "Approximating Correlated Equilibria using Relaxations on the Marginal Polytope", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Neural Information Processing Systems", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1153, + "page_form": 1153, "page_to": 1160, "totle_page": 8, "language": "en", - "abstract": "Abstract:In game theory, a Correlated Equilibrium (CE) is an equilibrium concept that generalizes the more well-known Nash Equilibrium. If the game is represented as a graphical game, the computational complexity of computing an optimum CE is exponential in the tree-width of the graph. In settings where this exact computation is not feasible, it is desirable to approximate the properties of the CE, such as its expected social utility and marginal probabilities. We study outer relaxations of this problem that yield approximate marginal strategies for the players under a variety of utility functions. Results on simulated games and in a real problem involving drug design indicate that our approximations can be highly accurate and can be successfully used when exact computation of CE is infeasible.\n" + "abstract": "Abstract:In game theory, a Correlated Equilibrium (CE) is an equilibrium concept that generalizes the more well-known Nash Equilibrium. If the game is represented as a graphical game, the computational complexity of computing an optimum CE is exponential in the tree-width of the graph. In settings where this exact computation is not feasible, it is desirable to approximate the properties of the CE, such as its expected social utility and marginal probabilities. We study outer relaxations of this problem that yield approximate marginal strategies for the players under a variety of utility functions. Results on simulated games and in a real problem involving drug design indicate that our approximations can be highly accurate and can be successfully used when exact computation of CE is infeasible." }, { "author": [ @@ -2321,14 +2321,14 @@ "Dy, Jennifer" ], "paper_title": "Active Learning from Crowds ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Machine Learning", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1161, + "page_form": 1161, "page_to": 1168, "totle_page": 8, "language": "en", - "abstract": "Abstract:Obtaining labels is expensive or time-consuming, but unlabeled data is often abundant and easy to obtain. Many learning task can profit from intelligently choosing unlabeled instances to be labeled by an oracle also known as active learning, instead of simply labeling all the data or randomly selecting data to be labeled. Supervised learning traditionally relies on an oracle playing the role of a teacher. In the multiple annotator paradigm, an oracle, who knows the ground truth, no longer exists; instead, multiple labelers, with varying expertise, are available for querying. This paradigm posits new challenges to the active learning scenario. We can ask which data sample should be labeled next and which annotator should we query to benefit our learning model the most. In this paper, we develop a probabilistic model for learning from multiple annotators that can also learn the annotator expertise even when their expertise may not be consistently accurate (or inaccurate) across the task domain. In addition, we provide an optimization fromulation that allows us to simultaneously learn the most uncertain sample and the annotator/s to query the labels from for active learning. Our active learning approach combines both intelligently selecting samples to label and learning from expertise among multiple labelers to improve learning perfromance.\n" + "abstract": "Abstract:Obtaining labels is expensive or time-consuming, but unlabeled data is often abundant and easy to obtain. Many learning task can profit from intelligently choosing unlabeled instances to be labeled by an oracle also known as active learning, instead of simply labeling all the data or randomly selecting data to be labeled. Supervised learning traditionally relies on an oracle playing the role of a teacher. In the multiple annotator paradigm, an oracle, who knows the ground truth, no longer exists; instead, multiple labelers, with varying expertise, are available for querying. This paradigm posits new challenges to the active learning scenario. We can ask which data sample should be labeled next and which annotator should we query to benefit our learning model the most. In this paper, we develop a probabilistic model for learning from multiple annotators that can also learn the annotator expertise even when their expertise may not be consistently accurate (or inaccurate) across the task domain. In addition, we provide an optimization formulation that allows us to simultaneously learn the most uncertain sample and the annotator/s to query the labels from for active learning. Our active learning approach combines both intelligently selecting samples to label and learning from expertise among multiple labelers to improve learning performance." }, { "author": [ @@ -2337,14 +2337,14 @@ "Bagnell, Drew" ], "paper_title": "Computational Rationalization: The Inverse Equilibrium Problem", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Uncertainty in Artificial Intelligence", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1169, + "page_form": 1169, "page_to": 1176, "totle_page": 8, "language": "en", - "abstract": "Abstract:Modeling the purposeful behavior of imperfect agents from a small number of observations is a challenging task. When restricted to the single-agent decision-theoretic setting, inverse optimal control techniques assume that observed behavior is an approximately optimal solution to an unknown decision problem. These techniques learn a utility function that explains the example behavior and can then be used to accurately predict or imitate future behavior in similar observed or unobserved situations. In this work, we consider similar tasks in competitive and cooperative multi-agent domains. Here, unlike single-agent settings, a player cannot myopically maximize its reward --- it must speculate on how the other agents may act to influence the game's outcome. Employing the game-theoretic notion of regret and the principle of maximum entropy, we introduce a technique for predicting and generalizing behavior, as well as recovering a reward function in these domains.\n" + "abstract": "Abstract:Modeling the purposeful behavior of imperfect agents from a small number of observations is a challenging task. When restricted to the single-agent decision-theoretic setting, inverse optimal control techniques assume that observed behavior is an approximately optimal solution to an unknown decision problem. These techniques learn a utility function that explains the example behavior and can then be used to accurately predict or imitate future behavior in similar observed or unobserved situations. In this work, we consider similar tasks in competitive and cooperative multi-agent domains. Here, unlike single-agent settings, a player cannot myopically maximize its reward --- it must speculate on how the other agents may act to influence the game's outcome. Employing the game-theoretic notion of regret and the principle of maximum entropy, we introduce a technique for predicting and generalizing behavior, as well as recovering a reward function in these domains." }, { "author": [ @@ -2354,14 +2354,14 @@ "Hoffman, Matthew" ], "paper_title": "Finite-Sample Analysis of Lasso-TD ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Intelligent RObots and IROS", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1177, + "page_form": 1177, "page_to": 1184, "totle_page": 8, "language": "en", - "abstract": "Abstract:In this paper, we analyze the perfromance of Lasso-TD, a modification of LSTD in which the projection operator is defined as a Lasso problem. We first show that Lasso-TD is guaranteed to have a unique fixed point and its algorithmic implementation coincides with the recently presented LARS-TD and LC-TD methods. We then derive two bounds on the prediction error of Lasso-TD in the Markov design setting, i.e., when the perfromance is evaluated on the same states used by the method. The first bound makes no assumption, but has a slow rate w.r.t. the number of samples. The second bound is under an assumption on the empirical Gram matrix, called the compatibility condition, but has an improved rate and directly relates the prediction error to the sparsity of the value function in the feature space at hand.\n" + "abstract": "Abstract:In this paper, we analyze the performance of Lasso-TD, a modification of LSTD in which the projection operator is defined as a Lasso problem. We first show that Lasso-TD is guaranteed to have a unique fixed point and its algorithmic implementation coincides with the recently presented LARS-TD and LC-TD methods. We then derive two bounds on the prediction error of Lasso-TD in the Markov design setting, i.e., when the performance is evaluated on the same states used by the method. The first bound makes no assumption, but has a slow rate w.r.t. the number of samples. The second bound is under an assumption on the empirical Gram matrix, called the compatibility condition, but has an improved rate and directly relates the prediction error to the sparsity of the value function in the feature space at hand." }, { "author": [ @@ -2369,14 +2369,14 @@ "Parr, Ron" ], "paper_title": "Generalized Value Functions for Large Action Sets", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Conference on Pattern Recognition", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1185, + "page_form": 1185, "page_to": 1192, "totle_page": 8, "language": "en", - "abstract": "Abstract:The majority of value function approximation based reinforcement learning algorithms available today, focus on approximating the state (V) or state-action (Q) value function and efficient action selection comes as an afterthought. On the other hand, real-world problems tend to have large action spaces, where evaluating every possible action becomes impractical. This mismatch presents a major obstacle in successfully applying reinforcement learning to real-world problems. In this paper we present a unified view of V and Q functions and arrive at a new space-efficient representation, where action selection can be done exponentially faster, without the use of a model. We then describe how to calculate this new value function efficiently via approximate linear programming and provide experimental results that demonstrate the effectiveness of the proposed approach.\n" + "abstract": "Abstract:The majority of value function approximation based reinforcement learning algorithms available today, focus on approximating the state (V) or state-action (Q) value function and efficient action selection comes as an afterthought. On the other hand, real-world problems tend to have large action spaces, where evaluating every possible action becomes impractical. This mismatch presents a major obstacle in successfully applying reinforcement learning to real-world problems. In this paper we present a unified view of V and Q functions and arrive at a new space-efficient representation, where action selection can be done exponentially faster, without the use of a model. We then describe how to calculate this new value function efficiently via approximate linear programming and provide experimental results that demonstrate the effectiveness of the proposed approach." }, { "author": [ @@ -2384,14 +2384,14 @@ "Taskar, Ben" ], "paper_title": "k-DPPs: Fixed-Size Determinantal Point Processes ", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "International Symposium on Neural Networks", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1193, + "page_form": 1193, "page_to": 1200, "totle_page": 8, "language": "en", - "abstract": "Abstract:We consider estimation methods for the class of continuous-data energy based models (EBMs). Our main result shows that estimating the parameters of an EBM using score matching when the conditional distribution over the visible units is Gaussian corresponds to training a particular from of regularized autoencoder. We show how different Gaussian EBMs lead to different autoencoder architectures, providing deep links between these two families of models. We compare the score matching estimator for the mPoT model, a particular Gaussian EBM, to several other training methods on a variety of tasks including image denoising and unsupervised feature extraction. We show that the regularization function induced by score matching leads to superior classification perfromance relative to a standard autoencoder. We also show that score matching yields classification results that are indistinguishable from better-known stochastic approximation maximum likelihood estimators.\n" + "abstract": "Abstract:We consider estimation methods for the class of continuous-data energy based models (EBMs). Our main result shows that estimating the parameters of an EBM using score matching when the conditional distribution over the visible units is Gaussian corresponds to training a particular form of regularized autoencoder. We show how different Gaussian EBMs lead to different autoencoder architectures, providing deep links between these two families of models. We compare the score matching estimator for the mPoT model, a particular Gaussian EBM, to several other training methods on a variety of tasks including image denoising and unsupervised feature extraction. We show that the regularization function induced by score matching leads to superior classification performance relative to a standard autoencoder. We also show that score matching yields classification results that are indistinguishable from better-known stochastic approximation maximum likelihood estimators." }, { "author": [ @@ -2402,14 +2402,14 @@ "Freitas, Nando" ], "paper_title": "On Autoencoders and Score Matching for Energy Based Models", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "Computational Learning Theory", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1201, + "page_form": 1201, "page_to": 1208, "totle_page": 8, "language": "en", - "abstract": "Abstract:Boosting is a popular way to derive powerful learners from simpler hypothesis classes. Following previous work (Mason et al., 1999; Friedman, 2000) on general boosting frameworks, we analyze gradient-based descent algorithms for boosting with respect to any convex objective and introduce a new measure of weak learner perfromance into this setting which generalizes existing work. We present the first weak to strong learning guarantees for the existing gradient boosting work for smooth convex objectives, and also demonstrate that this work fails for non-smooth objectives. To address this issue, we present new algorithms which extend this boosting approach to arbitrary convex loss functions and give corresponding weak to strong convergence results. In addition, we demonstrate experimental results that support our analysis and demonstrate the need for the new algorithms we present.\n" + "abstract": "Abstract:Boosting is a popular way to derive powerful learners from simpler hypothesis classes. Following previous work (Mason et al., 1999; Friedman, 2000) on general boosting frameworks, we analyze gradient-based descent algorithms for boosting with respect to any convex objective and introduce a new measure of weak learner performance into this setting which generalizes existing work. We present the first weak to strong learning guarantees for the existing gradient boosting work for smooth convex objectives, and also demonstrate that this work fails for non-smooth objectives. To address this issue, we present new algorithms which extend this boosting approach to arbitrary convex loss functions and give corresponding weak to strong convergence results. In addition, we demonstrate experimental results that support our analysis and demonstrate the need for the new algorithms we present." }, { "author": [ @@ -2417,13 +2417,13 @@ "Bagnell, Drew" ], "paper_title": "Generalized Boosting Algorithms for Convex Optimization", - "booktitle": "Proceedings of the 28th International Conference on Machine Learning (ICML-11)", + "booktitle": "The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases", "year": "2011", "isbn": "978-1-4503-0619-5", - "page_from": 1209, + "page_form": 1209, "page_to": 1216, "totle_page": 8, "language": "en", - "abstract": "Abstract:Determinantal point processes (DPPs) have recently been proposed as models for set selection problems where diversity is preferred. For example, they can be used to select diverse sets of sentences to from document summaries, or to find multiple non-overlapping human poses in an image. However, DPPs conflate the modeling of two distinct characteristics: the size of the set, and its content. For many realistic tasks, the size of the desired set is known up front; e.g., in search we may want to show the user exactly ten results. In these situations the effort spent by DPPs modeling set size is not only wasteful, but actually introduces unwanted bias into the modeling of content. Instead, we propose t\n" + "abstract": "Abstract:Determinantal point processes (DPPs) have recently been proposed as models for set selection problems where diversity is preferred. For example, they can be used to select diverse sets of sentences to form document summaries, or to find multiple non-overlapping human poses in an image. However, DPPs conflate the modeling of two distinct characteristics: the size of the set, and its content. For many realistic tasks, the size of the desired set is known up front; e.g., in search we may want to show the user exactly ten results. In these situations the effort spent by DPPs modeling set size is not only wasteful, but actually introduces unwanted bias into the modeling of content. Instead, we propose t" } ] diff --git a/db/seeds.rb b/db/seeds.rb index cf05956a..e28da95e 100644 --- a/db/seeds.rb +++ b/db/seeds.rb @@ -1,62 +1,87 @@ require 'factory_girl' +require 'faker' require 'json' data = File.read("db/data") data_json = JSON.parse(data) +name_tw = Array.new 51,"" +name_tw = name_tw.map do |p| Faker::Name::name end +name_en = Array.new 51,"" +name_en = name_en.map do |p| Faker::Name::name end + +email = Array.new 51,"" +email= email.map do |p| Faker::Internet.email end + +type = ["friend", "teacher", "student", "mate", "relation"] FactoryGirl.define do - factory :paper_record, class: "WritingJournal" do |f| - f.sequence(:paper_title_translations) do |n| - { zh_tw: "tw_#{data_json[n]["paper_title"]}", - en: "en_#{data_json[n]["paper_title"]}" } - end + factory(:co_author_candidate, class: "CoAuthor") do |f| - f.sequence(:journal_title_translations) do |n| - {zh_tw: "tw_#{data_json[n]["booktitle"]}", - en: "en_#{data_json[n]["booktitle"]}"} + f.sequence(:co_author_translations) do |n| + { zh_tw: "#{name_tw[n]}", + en: "#{name_en[n]}" } end - - f.sequence(:abstract) do |n| - "#{data_json[n]["abstract"]}" - end - - f.sequence(:isbn) do |n| - "#{data_json[n]["isbn"]}" - end - - f.sequence(:year) do |n| - "#{data_json[n]["year"]}" - end - - f.sequence(:authors) do |n| - "#{data_json[n]["author"].map{|m| m.split(",").reverse.join(" ")}.join(",")}" - end - - f.sequence(:form_to_start) do |n| - "#{data_json[n]["page_from"]}" - end - - f.sequence(:form_to_end) do |n| - "#{data_json[n]["page_to"]}" - end - - f.sequence(:total_pages) do |n| - "#{data_json[n]["total_page"]}" - end - - f.sequence(:language) do |n| - "#{data_json[n]["language"]}" - end - - f.sequence(:keywords) do |n| - "#{data_json[n]["abstract"].split[-3..-1].join(",")}" - end - - f.create_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account - f.update_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account - end + f.sequence(:type) do |n| "#{type[n%5]}" end + f.sequence(:email) do |n| "#{email[n]}" end + f.name_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account + end end +# factory :paper_record, class: "WritingJournal" do |f| +# f.sequence(:paper_title_translations) do |n| +# { zh_tw: "tw_#{data_json[n]["paper_title"]}", +# en: "en_#{data_json[n]["paper_title"]}" } +# end +# +# f.sequence(:journal_title_translations) do |n| +# {zh_tw: "tw_#{data_json[n]["booktitle"]}", +# en: "en_#{data_json[n]["booktitle"]}"} +# end +# +# f.sequence(:abstract) do |n| +# "#{data_json[n]["abstract"]}" +# end +# +# f.sequence(:isbn) do |n| +# "#{data_json[n]["isbn"]}" +# end +# +# f.sequence(:year) do |n| +# "#{data_json[n]["year"]}" +# end +# +# f.sequence(:authors) do |n| +# "#{data_json[n]["author"].map{|m| m.split(",").reverse.join(" ")}.join(",")}" +# end +# +# f.sequence(:form_to_start) do |n| +# "#{data_json[n]["page_from"]}" +# end +# +# f.sequence(:form_to_end) do |n| +# "#{data_json[n]["page_to"]}" +# end +# +# f.sequence(:total_pages) do |n| +# "#{data_json[n]["total_page"]}" +# end +# +# f.sequence(:language) do |n| +# "#{data_json[n]["language"]}" +# end +# +# f.sequence(:keywords) do |n| +# "#{data_json[n]["abstract"].split[-3..-1].join(",")}" +# end +# +# f.create_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account +# f.update_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account +# end + + +#50.times.each do +# FactoryGirl.create(:paper_record) +#end 50.times.each do - FactoryGirl.create(:paper_record) + FactoryGirl.create(:co_author_candidate) end From 46576c6947bf27944ac970aa05278b2ae85fad4a Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Tue, 27 Nov 2012 00:18:34 +0800 Subject: [PATCH 06/83] modify fake data --- db/seeds.rb | 107 ++++++++++++++++++++++++++-------------------------- 1 file changed, 54 insertions(+), 53 deletions(-) diff --git a/db/seeds.rb b/db/seeds.rb index e28da95e..3ecb2b43 100644 --- a/db/seeds.rb +++ b/db/seeds.rb @@ -25,62 +25,63 @@ FactoryGirl.define do f.sequence(:email) do |n| "#{email[n]}" end f.name_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account end + + factory :paper_record, class: "WritingJournal" do |f| + f.sequence(:paper_title_translations) do |n| + { zh_tw: "tw_#{data_json[n]["paper_title"]}", + en: "en_#{data_json[n]["paper_title"]}" } + end + + f.sequence(:journal_title_translations) do |n| + {zh_tw: "tw_#{data_json[n]["booktitle"]}", + en: "en_#{data_json[n]["booktitle"]}"} + end + + f.sequence(:abstract) do |n| + "#{data_json[n]["abstract"]}" + end + + f.sequence(:isbn) do |n| + "#{data_json[n]["isbn"]}" + end + + f.sequence(:year) do |n| + "#{data_json[n]["year"]}" + end + + f.sequence(:authors) do |n| + "#{data_json[n]["author"].map{|m| m.split(",").reverse.join(" ")}.join(",")}" + end + + f.sequence(:form_to_start) do |n| + "#{data_json[n]["page_from"]}" + end + + f.sequence(:form_to_end) do |n| + "#{data_json[n]["page_to"]}" + end + + f.sequence(:total_pages) do |n| + "#{data_json[n]["total_page"]}" + end + + f.sequence(:language) do |n| + "#{data_json[n]["language"]}" + end + + f.sequence(:keywords) do |n| + "#{data_json[n]["abstract"].split[-3..-1].join(",")}" + end + + f.create_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account + f.update_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account + end end -# factory :paper_record, class: "WritingJournal" do |f| -# f.sequence(:paper_title_translations) do |n| -# { zh_tw: "tw_#{data_json[n]["paper_title"]}", -# en: "en_#{data_json[n]["paper_title"]}" } -# end -# -# f.sequence(:journal_title_translations) do |n| -# {zh_tw: "tw_#{data_json[n]["booktitle"]}", -# en: "en_#{data_json[n]["booktitle"]}"} -# end -# -# f.sequence(:abstract) do |n| -# "#{data_json[n]["abstract"]}" -# end -# -# f.sequence(:isbn) do |n| -# "#{data_json[n]["isbn"]}" -# end -# -# f.sequence(:year) do |n| -# "#{data_json[n]["year"]}" -# end -# -# f.sequence(:authors) do |n| -# "#{data_json[n]["author"].map{|m| m.split(",").reverse.join(" ")}.join(",")}" -# end -# -# f.sequence(:form_to_start) do |n| -# "#{data_json[n]["page_from"]}" -# end -# -# f.sequence(:form_to_end) do |n| -# "#{data_json[n]["page_to"]}" -# end -# -# f.sequence(:total_pages) do |n| -# "#{data_json[n]["total_page"]}" -# end -# -# f.sequence(:language) do |n| -# "#{data_json[n]["language"]}" -# end -# -# f.sequence(:keywords) do |n| -# "#{data_json[n]["abstract"].split[-3..-1].join(",")}" -# end -# -# f.create_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account -# f.update_user_id BSON::ObjectId('4f45f3b9e9d02c5db9000067') #user_id, this is Chris' account -# end -#50.times.each do -# FactoryGirl.create(:paper_record) -#end +50.times.each do + FactoryGirl.create(:paper_record) +end 50.times.each do FactoryGirl.create(:co_author_candidate) From 9cc518b456d8b870aa4a752d2b6202462f7c1ef9 Mon Sep 17 00:00:00 2001 From: Harry Bomrah Date: Tue, 27 Nov 2012 15:55:32 +0800 Subject: [PATCH 07/83] autocomplete done --- .../javascripts/desktop/journal_pages.js.erb | 2 + app/assets/javascripts/orbitdesktop.js | 41 +++++++++++++++++++ app/assets/stylesheets/desktop.css | 1 + .../desktop/journal_pages/_form.html.erb | 13 ++++-- 4 files changed, 53 insertions(+), 4 deletions(-) diff --git a/app/assets/javascripts/desktop/journal_pages.js.erb b/app/assets/javascripts/desktop/journal_pages.js.erb index 9b7c0a03..f2cdab80 100644 --- a/app/assets/javascripts/desktop/journal_pages.js.erb +++ b/app/assets/javascripts/desktop/journal_pages.js.erb @@ -229,6 +229,8 @@ orbitDesktop.prototype.initializeJournalPapers = function(target,url,cache){ // }) } bindHandlers(); + + } this.initializeJournalPapers.coAuthorformCallback = function(data){ if(data.success){ diff --git a/app/assets/javascripts/orbitdesktop.js b/app/assets/javascripts/orbitdesktop.js index dc183d3d..a37154d9 100755 --- a/app/assets/javascripts/orbitdesktop.js +++ b/app/assets/javascripts/orbitdesktop.js @@ -6,6 +6,7 @@ //container=true is the area where the view will be loaded //load = true is used to load the list element by default //response-type = "json"|"script"|"xml|html" default is json +//autocomplete-list = "listname" an array from which autocomplete will be attached to its respective input or textarea $.extend($.expr[':'], { @@ -172,6 +173,7 @@ var orbitDesktop = function(dom){ }) } }); + var $widget_fn = $('.widget_fn'),$fn_des = $('.fn_des'); $widget_fn.hover(function(){ var fn_name = $(this).find('img').attr('alt'),nth = $(this).parents('.d_cate').index(),des_left = $('.dock_child').eq(nth).width(); @@ -181,6 +183,45 @@ var orbitDesktop = function(dom){ $(this).removeClass('thmc1'); $fn_des.stop(true, true).fadeOut(); }); + + var split = function( val ) { + return val.split( /,\s*/ ); + } + var extractLast = function( term ) { + return split( term ).pop(); + } + var autocompleteListName = null; + $("body").on("keydown","*[autocomplete-list]", function( event ) { + autocompleteListName = $(this).attr("autocomplete-list"); + if ( event.keyCode === $.ui.keyCode.TAB && + $( this ).data( "autocomplete" ).menu.active ) { + event.preventDefault(); + } + $(this).autocomplete({ + minLength: 0, + source: function( request, response ) { + // delegate back to autocomplete, but extract the last term + response( $.ui.autocomplete.filter( + window.o[o.data_method][autocompleteListName], extractLast( request.term ) ) ); + }, + focus: function() { + // prevent value inserted on focus + return false; + }, + select: function( event, ui ) { + var terms = split( this.value ); + // remove the current input + terms.pop(); + // add the selected item + terms.push( ui.item.value ); + // add placeholder to get the comma-and-space at the end + terms.push( "" ); + this.value = terms.join( ", " ); + return false; + } + }) + }) + }; this.sub_menu_item = function(dom){ if(!dom.hasClass('active')){ diff --git a/app/assets/stylesheets/desktop.css b/app/assets/stylesheets/desktop.css index 226afb8f..48c2ee6a 100644 --- a/app/assets/stylesheets/desktop.css +++ b/app/assets/stylesheets/desktop.css @@ -6,6 +6,7 @@ *= require bootstrap *= require bootstrap-orbit *= require jquery.miniColors + *= require jquery-ui *= require font-awesome *= require desktopmain *= require desktopmedia diff --git a/app/views/desktop/journal_pages/_form.html.erb b/app/views/desktop/journal_pages/_form.html.erb index 334a6eee..af7b97bf 100644 --- a/app/views/desktop/journal_pages/_form.html.erb +++ b/app/views/desktop/journal_pages/_form.html.erb @@ -52,7 +52,7 @@ <%= f.text_area :paper_title, size: "20x2", placeholder: "Paper Title", class: "s_grid_6 s_grid"%>
  • - <%= f.text_field :journal_title, size: "20", placeholder: "Journal Title", class: "s_grid_6 s_grid"%> + <%= f.text_field :journal_title, size: "20", placeholder: "Journal Title", class: "s_grid_6 s_grid", "autocomplete-list" => "journal_title_autocomplete_list" %>
  • @@ -113,7 +113,7 @@
    • - <%= f.text_area :authors, size: "20x2", placeholder: "Authors", class: "s_grid_6 s_grid"%> + <%= f.text_area :authors, size: "20x2", placeholder: "Authors", class: "s_grid_6 s_grid", "autocomplete-list" => "coauthor_autocomplete_list"%>
    • @@ -207,6 +207,11 @@
    - + + + + From 2f05885c4a13da7298c869115c0fb3bd04907774 Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 11:45:23 +0800 Subject: [PATCH 08/83] add journal list template --- .../desktop/journal_lists_controller.rb | 83 +++++++++ app/helpers/desktop/journal_lists_helper.rb | 2 + .../desktop/journal_lists/_form.html.erb | 17 ++ app/views/desktop/journal_lists/edit.html.erb | 6 + .../desktop/journal_lists/index.html.erb | 21 +++ app/views/desktop/journal_lists/new.html.erb | 5 + app/views/desktop/journal_lists/show.html.erb | 5 + .../desktop/journal_lists_controller_spec.rb | 164 ++++++++++++++++++ .../desktop/journal_lists_helper_spec.rb | 15 ++ .../desktop/desktop_journal_lists_spec.rb | 11 ++ .../desktop/journal_lists_routing_spec.rb | 35 ++++ .../journal_lists/edit.html.erb_spec.rb | 15 ++ .../journal_lists/index.html.erb_spec.rb | 15 ++ .../journal_lists/new.html.erb_spec.rb | 15 ++ .../journal_lists/show.html.erb_spec.rb | 12 ++ 15 files changed, 421 insertions(+) create mode 100644 app/controllers/desktop/journal_lists_controller.rb create mode 100644 app/helpers/desktop/journal_lists_helper.rb create mode 100644 app/views/desktop/journal_lists/_form.html.erb create mode 100644 app/views/desktop/journal_lists/edit.html.erb create mode 100644 app/views/desktop/journal_lists/index.html.erb create mode 100644 app/views/desktop/journal_lists/new.html.erb create mode 100644 app/views/desktop/journal_lists/show.html.erb create mode 100644 spec/controllers/desktop/journal_lists_controller_spec.rb create mode 100644 spec/helpers/desktop/journal_lists_helper_spec.rb create mode 100644 spec/requests/desktop/desktop_journal_lists_spec.rb create mode 100644 spec/routing/desktop/journal_lists_routing_spec.rb create mode 100644 spec/views/desktop/journal_lists/edit.html.erb_spec.rb create mode 100644 spec/views/desktop/journal_lists/index.html.erb_spec.rb create mode 100644 spec/views/desktop/journal_lists/new.html.erb_spec.rb create mode 100644 spec/views/desktop/journal_lists/show.html.erb_spec.rb diff --git a/app/controllers/desktop/journal_lists_controller.rb b/app/controllers/desktop/journal_lists_controller.rb new file mode 100644 index 00000000..f4a0d980 --- /dev/null +++ b/app/controllers/desktop/journal_lists_controller.rb @@ -0,0 +1,83 @@ +class Desktop::JournalListsController < ApplicationController + # GET /desktop/journal_lists + # GET /desktop/journal_lists.json + def index + @desktop_journal_lists = Desktop::JournalList.all + + respond_to do |format| + format.html # index.html.erb + format.json { render json: @desktop_journal_lists } + end + end + + # GET /desktop/journal_lists/1 + # GET /desktop/journal_lists/1.json + def show + @desktop_journal_list = Desktop::JournalList.find(params[:id]) + + respond_to do |format| + format.html # show.html.erb + format.json { render json: @desktop_journal_list } + end + end + + # GET /desktop/journal_lists/new + # GET /desktop/journal_lists/new.json + def new + @desktop_journal_list = Desktop::JournalList.new + + respond_to do |format| + format.html # new.html.erb + format.json { render json: @desktop_journal_list } + end + end + + # GET /desktop/journal_lists/1/edit + def edit + @desktop_journal_list = Desktop::JournalList.find(params[:id]) + end + + # POST /desktop/journal_lists + # POST /desktop/journal_lists.json + def create + @desktop_journal_list = Desktop::JournalList.new(params[:desktop_journal_list]) + + respond_to do |format| + if @desktop_journal_list.save + format.html { redirect_to @desktop_journal_list, notice: 'Journal list was successfully created.' } + format.json { render json: @desktop_journal_list, status: :created, location: @desktop_journal_list } + else + format.html { render action: "new" } + format.json { render json: @desktop_journal_list.errors, status: :unprocessable_entity } + end + end + end + + # PUT /desktop/journal_lists/1 + # PUT /desktop/journal_lists/1.json + def update + @desktop_journal_list = Desktop::JournalList.find(params[:id]) + + respond_to do |format| + if @desktop_journal_list.update_attributes(params[:desktop_journal_list]) + format.html { redirect_to @desktop_journal_list, notice: 'Journal list was successfully updated.' } + format.json { head :ok } + else + format.html { render action: "edit" } + format.json { render json: @desktop_journal_list.errors, status: :unprocessable_entity } + end + end + end + + # DELETE /desktop/journal_lists/1 + # DELETE /desktop/journal_lists/1.json + def destroy + @desktop_journal_list = Desktop::JournalList.find(params[:id]) + @desktop_journal_list.destroy + + respond_to do |format| + format.html { redirect_to desktop_journal_lists_url } + format.json { head :ok } + end + end +end diff --git a/app/helpers/desktop/journal_lists_helper.rb b/app/helpers/desktop/journal_lists_helper.rb new file mode 100644 index 00000000..e21dc09f --- /dev/null +++ b/app/helpers/desktop/journal_lists_helper.rb @@ -0,0 +1,2 @@ +module Desktop::JournalListsHelper +end diff --git a/app/views/desktop/journal_lists/_form.html.erb b/app/views/desktop/journal_lists/_form.html.erb new file mode 100644 index 00000000..ac7f7ac6 --- /dev/null +++ b/app/views/desktop/journal_lists/_form.html.erb @@ -0,0 +1,17 @@ +<%= form_for(@desktop_journal_list) do |f| %> + <% if @desktop_journal_list.errors.any? %> +
    +

    <%= pluralize(@desktop_journal_list.errors.count, "error") %> prohibited this desktop_journal_list from being saved:

    + +
      + <% @desktop_journal_list.errors.full_messages.each do |msg| %> +
    • <%= msg %>
    • + <% end %> +
    +
    + <% end %> + +
    + <%= f.submit %> +
    +<% end %> diff --git a/app/views/desktop/journal_lists/edit.html.erb b/app/views/desktop/journal_lists/edit.html.erb new file mode 100644 index 00000000..d628e895 --- /dev/null +++ b/app/views/desktop/journal_lists/edit.html.erb @@ -0,0 +1,6 @@ +

    Editing desktop_journal_list

    + +<%= render 'form' %> + +<%= link_to 'Show', @desktop_journal_list %> | +<%= link_to 'Back', desktop_journal_lists_path %> diff --git a/app/views/desktop/journal_lists/index.html.erb b/app/views/desktop/journal_lists/index.html.erb new file mode 100644 index 00000000..618d2ae1 --- /dev/null +++ b/app/views/desktop/journal_lists/index.html.erb @@ -0,0 +1,21 @@ +

    Listing desktop_journal_lists

    + + + + + + + + +<% @desktop_journal_lists.each do |desktop_journal_list| %> + + + + + +<% end %> +
    <%= link_to 'Show', desktop_journal_list %><%= link_to 'Edit', edit_desktop_journal_list_path(desktop_journal_list) %><%= link_to 'Destroy', desktop_journal_list, confirm: 'Are you sure?', method: :delete %>
    + +
    + +<%= link_to 'New Journal list', new_desktop_journal_list_path %> diff --git a/app/views/desktop/journal_lists/new.html.erb b/app/views/desktop/journal_lists/new.html.erb new file mode 100644 index 00000000..95017361 --- /dev/null +++ b/app/views/desktop/journal_lists/new.html.erb @@ -0,0 +1,5 @@ +

    New desktop_journal_list

    + +<%= render 'form' %> + +<%= link_to 'Back', desktop_journal_lists_path %> diff --git a/app/views/desktop/journal_lists/show.html.erb b/app/views/desktop/journal_lists/show.html.erb new file mode 100644 index 00000000..d454b78f --- /dev/null +++ b/app/views/desktop/journal_lists/show.html.erb @@ -0,0 +1,5 @@ +

    <%= notice %>

    + + +<%= link_to 'Edit', edit_desktop_journal_list_path(@desktop_journal_list) %> | +<%= link_to 'Back', desktop_journal_lists_path %> diff --git a/spec/controllers/desktop/journal_lists_controller_spec.rb b/spec/controllers/desktop/journal_lists_controller_spec.rb new file mode 100644 index 00000000..99d2e2dc --- /dev/null +++ b/spec/controllers/desktop/journal_lists_controller_spec.rb @@ -0,0 +1,164 @@ +require 'spec_helper' + +# This spec was generated by rspec-rails when you ran the scaffold generator. +# It demonstrates how one might use RSpec to specify the controller code that +# was generated by Rails when you ran the scaffold generator. +# +# It assumes that the implementation code is generated by the rails scaffold +# generator. If you are using any extension libraries to generate different +# controller code, this generated spec may or may not pass. +# +# It only uses APIs available in rails and/or rspec-rails. There are a number +# of tools you can use to make these specs even more expressive, but we're +# sticking to rails and rspec-rails APIs to keep things simple and stable. +# +# Compared to earlier versions of this generator, there is very limited use of +# stubs and message expectations in this spec. Stubs are only used when there +# is no simpler way to get a handle on the object needed for the example. +# Message expectations are only used when there is no simpler way to specify +# that an instance is receiving a specific message. + +describe Desktop::JournalListsController do + + # This should return the minimal set of attributes required to create a valid + # Desktop::JournalList. As you add validations to Desktop::JournalList, be sure to + # update the return value of this method accordingly. + def valid_attributes + {} + end + + # This should return the minimal set of values that should be in the session + # in order to pass any filters (e.g. authentication) defined in + # Desktop::JournalListsController. Be sure to keep this updated too. + def valid_session + {} + end + + describe "GET index" do + it "assigns all desktop_journal_lists as @desktop_journal_lists" do + journal_list = Desktop::JournalList.create! valid_attributes + get :index, {}, valid_session + assigns(:desktop_journal_lists).should eq([journal_list]) + end + end + + describe "GET show" do + it "assigns the requested desktop_journal_list as @desktop_journal_list" do + journal_list = Desktop::JournalList.create! valid_attributes + get :show, {:id => journal_list.to_param}, valid_session + assigns(:desktop_journal_list).should eq(journal_list) + end + end + + describe "GET new" do + it "assigns a new desktop_journal_list as @desktop_journal_list" do + get :new, {}, valid_session + assigns(:desktop_journal_list).should be_a_new(Desktop::JournalList) + end + end + + describe "GET edit" do + it "assigns the requested desktop_journal_list as @desktop_journal_list" do + journal_list = Desktop::JournalList.create! valid_attributes + get :edit, {:id => journal_list.to_param}, valid_session + assigns(:desktop_journal_list).should eq(journal_list) + end + end + + describe "POST create" do + describe "with valid params" do + it "creates a new Desktop::JournalList" do + expect { + post :create, {:desktop_journal_list => valid_attributes}, valid_session + }.to change(Desktop::JournalList, :count).by(1) + end + + it "assigns a newly created desktop_journal_list as @desktop_journal_list" do + post :create, {:desktop_journal_list => valid_attributes}, valid_session + assigns(:desktop_journal_list).should be_a(Desktop::JournalList) + assigns(:desktop_journal_list).should be_persisted + end + + it "redirects to the created desktop_journal_list" do + post :create, {:desktop_journal_list => valid_attributes}, valid_session + response.should redirect_to(Desktop::JournalList.last) + end + end + + describe "with invalid params" do + it "assigns a newly created but unsaved desktop_journal_list as @desktop_journal_list" do + # Trigger the behavior that occurs when invalid params are submitted + Desktop::JournalList.any_instance.stub(:save).and_return(false) + post :create, {:desktop_journal_list => {}}, valid_session + assigns(:desktop_journal_list).should be_a_new(Desktop::JournalList) + end + + it "re-renders the 'new' template" do + # Trigger the behavior that occurs when invalid params are submitted + Desktop::JournalList.any_instance.stub(:save).and_return(false) + post :create, {:desktop_journal_list => {}}, valid_session + response.should render_template("new") + end + end + end + + describe "PUT update" do + describe "with valid params" do + it "updates the requested desktop_journal_list" do + journal_list = Desktop::JournalList.create! valid_attributes + # Assuming there are no other desktop_journal_lists in the database, this + # specifies that the Desktop::JournalList created on the previous line + # receives the :update_attributes message with whatever params are + # submitted in the request. + Desktop::JournalList.any_instance.should_receive(:update_attributes).with({'these' => 'params'}) + put :update, {:id => journal_list.to_param, :desktop_journal_list => {'these' => 'params'}}, valid_session + end + + it "assigns the requested desktop_journal_list as @desktop_journal_list" do + journal_list = Desktop::JournalList.create! valid_attributes + put :update, {:id => journal_list.to_param, :desktop_journal_list => valid_attributes}, valid_session + assigns(:desktop_journal_list).should eq(journal_list) + end + + it "redirects to the desktop_journal_list" do + journal_list = Desktop::JournalList.create! valid_attributes + put :update, {:id => journal_list.to_param, :desktop_journal_list => valid_attributes}, valid_session + response.should redirect_to(journal_list) + end + end + + describe "with invalid params" do + it "assigns the desktop_journal_list as @desktop_journal_list" do + journal_list = Desktop::JournalList.create! valid_attributes + # Trigger the behavior that occurs when invalid params are submitted + Desktop::JournalList.any_instance.stub(:save).and_return(false) + put :update, {:id => journal_list.to_param, :desktop_journal_list => {}}, valid_session + assigns(:desktop_journal_list).should eq(journal_list) + end + + it "re-renders the 'edit' template" do + journal_list = Desktop::JournalList.create! valid_attributes + # Trigger the behavior that occurs when invalid params are submitted + Desktop::JournalList.any_instance.stub(:save).and_return(false) + put :update, {:id => journal_list.to_param, :desktop_journal_list => {}}, valid_session + response.should render_template("edit") + end + end + end + + describe "DELETE destroy" do + it "destroys the requested desktop_journal_list" do + journal_list = Desktop::JournalList.create! valid_attributes + expect { + delete :destroy, {:id => journal_list.to_param}, valid_session + }.to change(Desktop::JournalList, :count).by(-1) + end + + it "redirects to the desktop_journal_lists list" do + journal_list = Desktop::JournalList.create! valid_attributes + delete :destroy, {:id => journal_list.to_param}, valid_session + response.should redirect_to(desktop_journal_lists_url) + end + end + +end diff --git a/spec/helpers/desktop/journal_lists_helper_spec.rb b/spec/helpers/desktop/journal_lists_helper_spec.rb new file mode 100644 index 00000000..47b911a9 --- /dev/null +++ b/spec/helpers/desktop/journal_lists_helper_spec.rb @@ -0,0 +1,15 @@ +require 'spec_helper' + +# Specs in this file have access to a helper object that includes +# the Desktop::JournalListsHelper. For example: +# +# describe Desktop::JournalListsHelper do +# describe "string concat" do +# it "concats two strings with spaces" do +# helper.concat_strings("this","that").should == "this that" +# end +# end +# end +describe Desktop::JournalListsHelper do + pending "add some examples to (or delete) #{__FILE__}" +end diff --git a/spec/requests/desktop/desktop_journal_lists_spec.rb b/spec/requests/desktop/desktop_journal_lists_spec.rb new file mode 100644 index 00000000..c1ac52e3 --- /dev/null +++ b/spec/requests/desktop/desktop_journal_lists_spec.rb @@ -0,0 +1,11 @@ +require 'spec_helper' + +describe "Desktop::JournalLists" do + describe "GET /desktop_journal_lists" do + it "works! (now write some real specs)" do + # Run the generator again with the --webrat flag if you want to use webrat methods/matchers + get desktop_journal_lists_path + response.status.should be(200) + end + end +end diff --git a/spec/routing/desktop/journal_lists_routing_spec.rb b/spec/routing/desktop/journal_lists_routing_spec.rb new file mode 100644 index 00000000..79eac59e --- /dev/null +++ b/spec/routing/desktop/journal_lists_routing_spec.rb @@ -0,0 +1,35 @@ +require "spec_helper" + +describe Desktop::JournalListsController do + describe "routing" do + + it "routes to #index" do + get("/desktop/journal_lists").should route_to("desktop/journal_lists#index") + end + + it "routes to #new" do + get("/desktop/journal_lists/new").should route_to("desktop/journal_lists#new") + end + + it "routes to #show" do + get("/desktop/journal_lists/1").should route_to("desktop/journal_lists#show", :id => "1") + end + + it "routes to #edit" do + get("/desktop/journal_lists/1/edit").should route_to("desktop/journal_lists#edit", :id => "1") + end + + it "routes to #create" do + post("/desktop/journal_lists").should route_to("desktop/journal_lists#create") + end + + it "routes to #update" do + put("/desktop/journal_lists/1").should route_to("desktop/journal_lists#update", :id => "1") + end + + it "routes to #destroy" do + delete("/desktop/journal_lists/1").should route_to("desktop/journal_lists#destroy", :id => "1") + end + + end +end diff --git a/spec/views/desktop/journal_lists/edit.html.erb_spec.rb b/spec/views/desktop/journal_lists/edit.html.erb_spec.rb new file mode 100644 index 00000000..92a7ab73 --- /dev/null +++ b/spec/views/desktop/journal_lists/edit.html.erb_spec.rb @@ -0,0 +1,15 @@ +require 'spec_helper' + +describe "desktop/journal_lists/edit" do + before(:each) do + @desktop_journal_list = assign(:desktop_journal_list, stub_model(Desktop::JournalList)) + end + + it "renders the edit desktop_journal_list form" do + render + + # Run the generator again with the --webrat flag if you want to use webrat matchers + assert_select "form", :action => desktop_journal_lists_path(@desktop_journal_list), :method => "post" do + end + end +end diff --git a/spec/views/desktop/journal_lists/index.html.erb_spec.rb b/spec/views/desktop/journal_lists/index.html.erb_spec.rb new file mode 100644 index 00000000..93bfc83b --- /dev/null +++ b/spec/views/desktop/journal_lists/index.html.erb_spec.rb @@ -0,0 +1,15 @@ +require 'spec_helper' + +describe "desktop/journal_lists/index" do + before(:each) do + assign(:desktop_journal_lists, [ + stub_model(Desktop::JournalList), + stub_model(Desktop::JournalList) + ]) + end + + it "renders a list of desktop/journal_lists" do + render + # Run the generator again with the --webrat flag if you want to use webrat matchers + end +end diff --git a/spec/views/desktop/journal_lists/new.html.erb_spec.rb b/spec/views/desktop/journal_lists/new.html.erb_spec.rb new file mode 100644 index 00000000..b9751d44 --- /dev/null +++ b/spec/views/desktop/journal_lists/new.html.erb_spec.rb @@ -0,0 +1,15 @@ +require 'spec_helper' + +describe "desktop/journal_lists/new" do + before(:each) do + assign(:desktop_journal_list, stub_model(Desktop::JournalList).as_new_record) + end + + it "renders new desktop_journal_list form" do + render + + # Run the generator again with the --webrat flag if you want to use webrat matchers + assert_select "form", :action => desktop_journal_lists_path, :method => "post" do + end + end +end diff --git a/spec/views/desktop/journal_lists/show.html.erb_spec.rb b/spec/views/desktop/journal_lists/show.html.erb_spec.rb new file mode 100644 index 00000000..d18b4ece --- /dev/null +++ b/spec/views/desktop/journal_lists/show.html.erb_spec.rb @@ -0,0 +1,12 @@ +require 'spec_helper' + +describe "desktop/journal_lists/show" do + before(:each) do + @desktop_journal_list = assign(:desktop_journal_list, stub_model(Desktop::JournalList)) + end + + it "renders attributes in

    " do + render + # Run the generator again with the --webrat flag if you want to use webrat matchers + end +end From 114ca5493acfc5e042ccba1c8c996afd023252cb Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 13:43:48 +0800 Subject: [PATCH 09/83] replace rss format with json in announcement --- .../back_end/bulletin_categorys_controller.rb | 13 +++++-- .../front_end/bulletins_controller.rb | 35 ++++++++++--------- .../front_end/bulletins/index.rss.builder | 14 ++++---- .../announcement/config/routes.rb | 6 ++-- 4 files changed, 40 insertions(+), 28 deletions(-) diff --git a/vendor/built_in_modules/announcement/app/controllers/panel/announcement/back_end/bulletin_categorys_controller.rb b/vendor/built_in_modules/announcement/app/controllers/panel/announcement/back_end/bulletin_categorys_controller.rb index 0151d4c0..1802db62 100644 --- a/vendor/built_in_modules/announcement/app/controllers/panel/announcement/back_end/bulletin_categorys_controller.rb +++ b/vendor/built_in_modules/announcement/app/controllers/panel/announcement/back_end/bulletin_categorys_controller.rb @@ -1,6 +1,7 @@ class Panel::Announcement::BackEnd::BulletinCategorysController < OrbitBackendController include OrbitControllerLib::DivisionForDisable - before_filter :for_app_manager,:except => [:index,:get_categorys_json,:get_bulletins_json] + # if someone want to use json format replace with rss, please add :get_bulletins_json to below + before_filter :for_app_manager,:except => [:index,:get_categorys_json] def index @bulletin_categorys = get_categories_for_index("BulletinCategory") @@ -22,7 +23,15 @@ class Panel::Announcement::BackEnd::BulletinCategorysController < OrbitBackendCo categorys.each do |c| data << { category: c.title, - link: "http://#{request.host_with_port}#{panel_announcement_back_end_bulletin_category_get_bulletins_json_path(c)}" + # this is a json format + # link: "http://#{request.host_with_port}#{panel_announcement_back_end_bulletin_category_get_bulletins_json_path(c)}" + # remember to uncommand a rule in route.rb, too + link: "#{url_for( :action => "index", + :controller => "panel/announcement/front_end/bulletins", + :format => :rss, + :only_path => false, + :inner=>true, + :category_id => c )}" } end diff --git a/vendor/built_in_modules/announcement/app/controllers/panel/announcement/front_end/bulletins_controller.rb b/vendor/built_in_modules/announcement/app/controllers/panel/announcement/front_end/bulletins_controller.rb index 8fa7af68..456d57d9 100644 --- a/vendor/built_in_modules/announcement/app/controllers/panel/announcement/front_end/bulletins_controller.rb +++ b/vendor/built_in_modules/announcement/app/controllers/panel/announcement/front_end/bulletins_controller.rb @@ -1,10 +1,10 @@ class Panel::Announcement::FrontEnd::BulletinsController < OrbitWidgetController - + def initialize super @app_title = 'announcement' end - + # GET /bulletins # GET /bulletins.xml def index_bulletins_by_unit @@ -14,14 +14,15 @@ class Panel::Announcement::FrontEnd::BulletinsController < OrbitWidgetController end def index - @item = Page.find(params[:page_id]) - - if @item.frontend_data_count - @page_num = @item.frontend_data_count - else - @page_num = 15 - end - @frontend_style = @item.frontend_style + @item = Page.find(params[:page_id]) rescue nil + if @item + if @item.frontend_data_count + @page_num = @item.frontend_data_count + else + @page_num = 15 + end + @frontend_style = @item.frontend_style + end @page = Page.find(params[:page_id]) rescue nil if !params[:search_query].blank? @@ -54,9 +55,9 @@ class Panel::Announcement::FrontEnd::BulletinsController < OrbitWidgetController end impressionist(@tag) if @tag end - + end - + def show @page = Page.find(params[:page_id]) rescue nil if params[:preview] == "true" @@ -71,22 +72,22 @@ class Panel::Announcement::FrontEnd::BulletinsController < OrbitWidgetController render :text => "

    #{t('sys.can_not_display_due_to_no_context')}
    ".html_safe end else - render :nothing => true, :status => 403 + render :nothing => true, :status => 403 end end end - + def preview_content @bulletin = Bulletin.find params[:id] rescue nil @bulletin = Preview.find(params[:id]).get_virtual_object if @bulletin.nil? get_categorys end - + protected - + def get_categorys @bulletin_categorys = BulletinCategory.all end - + end diff --git a/vendor/built_in_modules/announcement/app/views/panel/announcement/front_end/bulletins/index.rss.builder b/vendor/built_in_modules/announcement/app/views/panel/announcement/front_end/bulletins/index.rss.builder index ca4719c6..40698bfc 100644 --- a/vendor/built_in_modules/announcement/app/views/panel/announcement/front_end/bulletins/index.rss.builder +++ b/vendor/built_in_modules/announcement/app/views/panel/announcement/front_end/bulletins/index.rss.builder @@ -1,13 +1,13 @@ xml.instruct! :xml, :version => "1.0" xml.rss :version => "2.0" do xml.channel do - if @current_category - xml.title @current_category.title_translations[I18n.locale.to_s] - else - xml.title t('announcement.announcement') - end + if @current_category + xml.title @current_category.title_translations[I18n.locale.to_s] + else + xml.title t('announcement.announcement') + end xml.link url_for(:action=>"index", :controller=>"panel/announcement/front_end/bulletins",:format=> :rss,:only_path=>false,:inner=>true) - + for bulletin in @bulletins xml.item do xml.title bulletin.title_translations[I18n.locale.to_s] @@ -18,4 +18,4 @@ xml.rss :version => "2.0" do end end end -end \ No newline at end of file +end diff --git a/vendor/built_in_modules/announcement/config/routes.rb b/vendor/built_in_modules/announcement/config/routes.rb index 35500ac3..8c11ebd0 100644 --- a/vendor/built_in_modules/announcement/config/routes.rb +++ b/vendor/built_in_modules/announcement/config/routes.rb @@ -6,7 +6,7 @@ Rails.application.routes.draw do match 'approval_setting' => "approvals#setting" ,:as => :approval_setting,:via => :get match 'approval_setting' => "approvals#update_setting" ,:as => :approval_setting,:via => :post match 'approval_setting' => "approvals#user_list" ,:as => :approval_user_list,:via => :put - #match 'get_bulletins_json' => "bulletins#get_bulletins_json" ,:as => :bulletins_json_list,:via => :get + resources :bulletins do match "approve/:bulletin_id" => "approvals#preview_and_approve",:as => :approval_preview,:via => :put match "approve/:bulletin_id" => "approvals#approve",:as => :approve,:via => :post @@ -29,7 +29,9 @@ Rails.application.routes.draw do collection do get 'get_categorys_json' end - match "get_bulletins_json/" => "bulletin_categorys#get_bulletins_json", :as => :get_bulletins_json + # if want to use json + # please uncommand below line + # match "get_bulletins_json/" => "bulletin_categorys#get_bulletins_json", :as => :get_bulletins_json end resources :bulletin_links, :controller => 'bulletin_links' do From 730c29bf490fb35acc5408e1603f50f01416b43c Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 14:28:41 +0800 Subject: [PATCH 10/83] change layout --- config/application.rb | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/config/application.rb b/config/application.rb index f0e40fcc..0898a2e6 100644 --- a/config/application.rb +++ b/config/application.rb @@ -68,6 +68,10 @@ module Orbit config.assets.enabled = true #config.time_zone = 'Taipei' ENV['TZ'] = 'Asia/Taipei' + + config.to_prepare do + Devise::RegistrationsController.layout false + end end end Orbit_Apps = [] From 25324809bbd2223c236753241faf3d245ad4f0fa Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 14:42:15 +0800 Subject: [PATCH 11/83] add account link --- app/views/desktop/settings.html.erb | 48 ++++++++++++++--------------- 1 file changed, 24 insertions(+), 24 deletions(-) diff --git a/app/views/desktop/settings.html.erb b/app/views/desktop/settings.html.erb index df754325..88b948e5 100644 --- a/app/views/desktop/settings.html.erb +++ b/app/views/desktop/settings.html.erb @@ -1,26 +1,26 @@
    -
    - -
    -
    - -
    -
    - -
    -
    -
    +
    + +
    +
    + +
    +
    + +
    +
    +
    +
    -
    \ No newline at end of file From 796dcc4dc62198286409024a0fee958dd2a65e6d Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 15:20:02 +0800 Subject: [PATCH 12/83] easy to make fake data --- db/seeds.rb | 3 +++ 1 file changed, 3 insertions(+) diff --git a/db/seeds.rb b/db/seeds.rb index 3ecb2b43..00825e32 100644 --- a/db/seeds.rb +++ b/db/seeds.rb @@ -2,6 +2,9 @@ require 'factory_girl' require 'faker' require 'json' +WritingJournal.destroy_all +CoAuthor.destroy_all + data = File.read("db/data") data_json = JSON.parse(data) From 2f67a9653223ebeea3e4ffcfe8b34946bf32a92e Mon Sep 17 00:00:00 2001 From: Harry Bomrah Date: Wed, 28 Nov 2012 18:11:49 +0800 Subject: [PATCH 13/83] url fixed and edit form fixed for tinyscrollbar --- .../javascripts/desktop/journal_pages.js.erb | 4 + app/assets/javascripts/orbitdesktop.js | 109 ++++++++++++------ app/views/desktop/index.html.erb | 4 +- app/views/desktop/journal_p.html.erb | 2 +- app/views/desktop/settings.html.erb | 4 +- app/views/devise/registrations/edit.html.erb | 9 +- 6 files changed, 88 insertions(+), 44 deletions(-) diff --git a/app/assets/javascripts/desktop/journal_pages.js.erb b/app/assets/javascripts/desktop/journal_pages.js.erb index f2cdab80..44d3bc9c 100644 --- a/app/assets/javascripts/desktop/journal_pages.js.erb +++ b/app/assets/javascripts/desktop/journal_pages.js.erb @@ -57,6 +57,10 @@ orbitDesktop.prototype.initializeJournalPapers = function(target,url,cache){ // success : function(data){ var prev_data = $("div[container=true]").html(); $("div[container=true]").html(data); + o.tinyscrollbar_ext({ + main: '.tinycanvas', + fill: '.s_grid_con' + }) $(".bt-cancel").click(function(){ $("div[container=true]").html(prev_data); }) diff --git a/app/assets/javascripts/orbitdesktop.js b/app/assets/javascripts/orbitdesktop.js index a37154d9..f66eacb4 100755 --- a/app/assets/javascripts/orbitdesktop.js +++ b/app/assets/javascripts/orbitdesktop.js @@ -65,15 +65,28 @@ var orbitDesktop = function(dom){ o.loadWallpaper(customwallpaper); o.bindDesktopEvents(); o.loadIconCache(); - - $(o.contentHolder).empty().load("/desktop/desktop",function(){ - o.desktopData["d_desktop"] = ""; - o.initializeDesktop("d_desktop","",false); - }) + var custom_load = window.location.hash; + if(!custom_load){ + $(o.contentHolder).empty().load("/desktop/desktop",function(){ + o.desktopData["d_desktop"] = ""; + o.initializeDesktop("d_desktop","",false); + }) + }else{ + o.customPage(custom_load); + } }); }) } } + this.customPage = function(customload){ + customload = customload.replace("#",""); + if(customload.search("-") != -1){ + customload = customload.split("-"); + o.menu_item($(".docklist a[custom-load="+customload[0]+"]"),true,customload[1]); + }else{ + o.menu_item($(".docklist a[custom-load="+customload+"]"),true); + } + } this.changeTheme = function(theme){ // this function is used for changing theme o.theme = theme; $.getJSON("/"+o.themefolder+"/"+theme+"/settings/"+theme+".json",function(ts){ @@ -84,38 +97,10 @@ var orbitDesktop = function(dom){ o.loadIconCache(); }) }; + this.bindDesktopEvents = function(){ //this function will bind the global handlers to thd desktop, for example doc $(".docklist a").click(function(){ - var target = $(this).attr("id"); - var url = $(this).attr("href"); - o.data_method = $(this).attr("callback-method"); - if(o.currenthtml!=target){ - if(o.desktopData[o.currentface] == "undefined") - o.desktopData[o.currentface] = ""; - o.desktopData[o.currentface] = $(o.contentHolder).html(); - $("#content").hide("drop",o.transitionTime,function(){ - o.currenthtml = target; - o.currentface = target; - var cache = false; - if(!o.desktopData[o.currentface]){ - $(o.contentHolder).empty().load(url,function(){ - if(typeof o.data_method != "undefined"){ - if(o.data_method != "") - window.o[o.data_method](target,url,cache); - } - o.sub_menu_item($(o.contentHolder).find("*[content-type=menu] a[load=true]")); - }) - }else{ - $(o.contentHolder).html(o.desktopData[o.currentface]); - o.sub_menu_item($(o.contentHolder).find("*[content-type=menu] a[load=true]")); - cache = true; - if(typeof o.data_method != "undefined"){ - if(o.data_method != "") - window.o[o.data_method](target,url,cache); - } - } - }); - } + o.menu_item($(this)); return false; }) @@ -223,6 +208,58 @@ var orbitDesktop = function(dom){ }) }; + this.menu_item = function(dom,customload,submenuitem){ + if(!customload)customload=false; + var target = dom.attr("id"); + var url = dom.attr("href"); + o.data_method = dom.attr("callback-method"); + if(o.currenthtml!=target){ + if(o.desktopData[o.currentface] == "undefined") + o.desktopData[o.currentface] = ""; + o.desktopData[o.currentface] = $(o.contentHolder).html(); + if(customload){ + $(o.contentHolder).html("
    "); + } + $("#content").hide("drop",o.transitionTime,function(){ + o.currenthtml = target; + o.currentface = target; + + var cache = false; + if(!o.desktopData[o.currentface]){ + $(o.contentHolder).empty().load(url,function(){ + if(typeof o.data_method != "undefined"){ + if(o.data_method != "") + window.o[o.data_method](target,url,cache); + } + if(!customload) + o.sub_menu_item($(o.contentHolder).find("*[content-type=menu] a[load=true]")); + else{ + if(submenuitem) + o.sub_menu_item($(o.contentHolder).find("*[content-type=menu] a[custom-load="+submenuitem+"]")); + else + o.sub_menu_item($(o.contentHolder).find("*[content-type=menu] a[load=true]")); + } + + }) + }else{ + $(o.contentHolder).html(o.desktopData[o.currentface]); + if(!customload) + o.sub_menu_item($(o.contentHolder).find("*[content-type=menu] a[load=true]")); + else{ + if(submenuitem) + o.sub_menu_item($(o.contentHolder).find("*[content-type=menu] a[custom-load="+submenuitem+"]")); + else + o.sub_menu_item($(o.contentHolder).find("*[content-type=menu] a[load=true]")); + } + cache = true; + if(typeof o.data_method != "undefined"){ + if(o.data_method != "") + window.o[o.data_method](target,url,cache); + } + } + }); + } + } this.sub_menu_item = function(dom){ if(!dom.hasClass('active')){ var sub_data_method = dom.attr('callback-method'); @@ -623,6 +660,8 @@ var orbitDesktop = function(dom){ }; this.initializeSettings = function(target,url,cache){ //this is to initialize setting page + this.initializeSettings.account = function(){} + this.initializeSettings.sections = function(){ // this load section page in setting page var bindHandlers = function(){ // binding handlers in section page $('.tinycanvas').each(function(){ diff --git a/app/views/desktop/index.html.erb b/app/views/desktop/index.html.erb index b3f78780..c90f53fd 100755 --- a/app/views/desktop/index.html.erb +++ b/app/views/desktop/index.html.erb @@ -5,13 +5,13 @@
    • App Manager
    • All Sections
    • -
    • Settings
    • +
    • Settings
  • Publication
      -
    • Journal Papers
    • +
    • Journal Papers
    • Seminar Papers
    • diff --git a/app/views/desktop/journal_p.html.erb b/app/views/desktop/journal_p.html.erb index 87cf8a17..494e0001 100644 --- a/app/views/desktop/journal_p.html.erb +++ b/app/views/desktop/journal_p.html.erb @@ -14,7 +14,7 @@
      • List
      • -
      • Add/Edit
      • +
      • Add/Edit
      • Journals
      • Co-Authors
      • Tags & Keywords
      • diff --git a/app/views/desktop/settings.html.erb b/app/views/desktop/settings.html.erb index 88b948e5..590f4e89 100644 --- a/app/views/desktop/settings.html.erb +++ b/app/views/desktop/settings.html.erb @@ -10,8 +10,8 @@
        diff --git a/app/views/devise/registrations/edit.html.erb b/app/views/devise/registrations/edit.html.erb index ff380509..5f272075 100644 --- a/app/views/devise/registrations/edit.html.erb +++ b/app/views/devise/registrations/edit.html.erb @@ -1,6 +1,6 @@

        Edit <%= resource_name.to_s.humanize %>

        -<%= form_for(resource, :as => resource_name, :url => registration_path(resource_name), :html => { :method => :put }) do |f| %> +<%= form_for(resource, :as => resource_name, :url => registration_path(resource_name), :html => { :method => :put, "form-type"=>"ajax_form" }) do |f| %> <%= devise_error_messages! %>
        <%= f.label :email %>
        @@ -18,8 +18,9 @@
        <%= f.submit "Update" %>
        <% end %> -

        Cancel my account

        + \ No newline at end of file From 254a81fa5d4b089f56ce1ecaad9bcea4ae96fd39 Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 18:12:21 +0800 Subject: [PATCH 14/83] add templete --- .../desktop/journal_pages/_form.html.erb | 45 +++++++++++++++++++ 1 file changed, 45 insertions(+) diff --git a/app/views/desktop/journal_pages/_form.html.erb b/app/views/desktop/journal_pages/_form.html.erb index af7b97bf..e72db279 100644 --- a/app/views/desktop/journal_pages/_form.html.erb +++ b/app/views/desktop/journal_pages/_form.html.erb @@ -191,7 +191,35 @@ <%= wjf.file_field :file %> <% end %> --> + + + + + + + + + + + + + + + <% @writing_journal.writing_journal_files.each_with_index do |writing_journal_file, i| %> + <%= f.fields_for :writing_journal_files, writing_journal_file do |f| %> + <%= render :partial => 'form_file', :object => writing_journal_file, :locals => {:f => f, :i => i} %> + <% end %> + <% end %> + +
        FileFile Name
        +
        + <%= hidden_field_tag 'plugin_file_field_count', @writing_journal.writing_journal_files.count %> + add +
        +
        +
        +
      • <%= f.text_area :note, size: "20x2", plcaeholder: "Note", class: "s_grid_6 s_grid"%>
      • @@ -213,5 +241,22 @@ +<%= javascript_include_tag "archive_plugin_form" %> + From 6d159a3a9ef9b6484b15bc28f29541f80ad6545f Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 18:12:37 +0800 Subject: [PATCH 15/83] add templete --- .../desktop/journal_pages/_form_file.html.erb | 41 +++++++++++++++++++ 1 file changed, 41 insertions(+) create mode 100644 app/views/desktop/journal_pages/_form_file.html.erb diff --git a/app/views/desktop/journal_pages/_form_file.html.erb b/app/views/desktop/journal_pages/_form_file.html.erb new file mode 100644 index 00000000..b034589a --- /dev/null +++ b/app/views/desktop/journal_pages/_form_file.html.erb @@ -0,0 +1,41 @@ + +" class='list_item'> + +
        +
        + <%= f.file_field :file %> + <%= form_file.file.file ? ( link_to t(:view), form_file.file.url, {:class => 'btn', :target => '_blank', :title => t(:view)} ) : '' %> +
        +
        + + + +
        + <% @site_valid_locales.each_with_index do |locale, i| %> +
        "> + <%= f.fields_for :file_title_translations do |f| %> +
        + +
        + <%= f.text_field locale, :class=>'post-file_title', :value => (form_file.file_title_translations[locale] rescue nil) %> +
        +
        + <% end %> +
        + <% end %> +
        + + + + + <% if form_file.new_record? %> + + <% else %> + <%= f.hidden_field :id %> + + <%= f.hidden_field :should_destroy, :value => nil, :class => 'should_destroy' %> + <% end %> + + + + From 6ed1dd0cc6e62d81e6154636d1736028a28b5319 Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 23:43:39 +0800 Subject: [PATCH 16/83] change template --- app/views/desktop/journal_p.html.erb | 2 +- .../desktop/journal_pages/_form_file.html.erb | 52 +++++++++---------- 2 files changed, 26 insertions(+), 28 deletions(-) diff --git a/app/views/desktop/journal_p.html.erb b/app/views/desktop/journal_p.html.erb index 494e0001..b27c3560 100644 --- a/app/views/desktop/journal_p.html.erb +++ b/app/views/desktop/journal_p.html.erb @@ -14,7 +14,7 @@
        • List
        • -
        • Add/Edit
        • +
        • Add/Edit
        • Journals
        • Co-Authors
        • Tags & Keywords
        • diff --git a/app/views/desktop/journal_pages/_form_file.html.erb b/app/views/desktop/journal_pages/_form_file.html.erb index b034589a..29d21664 100644 --- a/app/views/desktop/journal_pages/_form_file.html.erb +++ b/app/views/desktop/journal_pages/_form_file.html.erb @@ -1,10 +1,9 @@ - " class='list_item'>
          - <%= f.file_field :file %> - <%= form_file.file.file ? ( link_to t(:view), form_file.file.url, {:class => 'btn', :target => '_blank', :title => t(:view)} ) : '' %> + <%= f.file_field :file, class: "s_grid_1 s_grid" %> + <%= form_file.file.file ? ( link_to t(:view), form_file.file.url, {:class => 'btn s_grid_1 s_grid', :target => '_blank', :title => t(:view)} ) : '' %>
          @@ -13,29 +12,28 @@
          <% @site_valid_locales.each_with_index do |locale, i| %>
          "> - <%= f.fields_for :file_title_translations do |f| %> -
          - -
          - <%= f.text_field locale, :class=>'post-file_title', :value => (form_file.file_title_translations[locale] rescue nil) %> -
          + <%= f.fields_for :file_title_translations do |f| %> +
          + +
          + <%= f.text_field locale, :class=>'s_grid_1 s_grid', :value => (form_file.file_title_translations[locale] rescue nil) %>
          - <% end %> -
          - <% end %> -
          - - - - - <% if form_file.new_record? %> - - <% else %> - <%= f.hidden_field :id %> - - <%= f.hidden_field :should_destroy, :value => nil, :class => 'should_destroy' %> - <% end %> - - - +
          + <% end %> +
          + <% end %> +
        + + + + <% if form_file.new_record? %> + + <% else %> + <%= f.hidden_field :id %> + + <%= f.hidden_field :should_destroy, :value => nil, :class => 'should_destroy' %> + <% end %> + + + From 9e2d6cfcdd0676f62426f2324c6fb55226b2698d Mon Sep 17 00:00:00 2001 From: Rueshyna Date: Wed, 28 Nov 2012 23:44:43 +0800 Subject: [PATCH 17/83] add quick link to change user password, but it's not work --- app/views/layouts/_orbit_bar.html.erb | 200 +++++++++++++------------- 1 file changed, 99 insertions(+), 101 deletions(-) diff --git a/app/views/layouts/_orbit_bar.html.erb b/app/views/layouts/_orbit_bar.html.erb index b3967885..224924c7 100644 --- a/app/views/layouts/_orbit_bar.html.erb +++ b/app/views/layouts/_orbit_bar.html.erb @@ -1,105 +1,103 @@ \ No newline at end of file + +
      • <%= t(:or_lower) %>
      • + <% end %> +
      + <%#= link_to t(:register), new_user_registration_path, :class => 'btn btn-danger register' %> +
      + + <% end %> +
    + + + From 390da0e49a603f218179e0da89bf8ded14c42eee Mon Sep 17 00:00:00 2001 From: chris Date: Thu, 29 Nov 2012 00:26:36 +0800 Subject: [PATCH 18/83] Change icons --- app/assets/fonts/entypo.eot | Bin 0 -> 24636 bytes app/assets/fonts/entypo.svg | 580 ++++++++++++ app/assets/fonts/entypo.ttf | Bin 0 -> 24476 bytes app/assets/fonts/entypo.woff | Bin 0 -> 32144 bytes app/assets/stylesheets/icons.css.erb | 519 +++++++++++ app/assets/stylesheets/new_admin.css.erb | 1 + app/assets/stylesheets/orbit-bar.css.erb | 16 +- app/assets/stylesheets/style.css.erb | 856 +----------------- app/helpers/application_helper.rb | 1 + app/views/admin/ad_images/_form.html.erb | 10 +- app/views/admin/assets/_asset.html.erb | 2 +- app/views/admin/dashboards/index.html.erb | 8 +- app/views/admin/plugins/index.html.erb | 2 +- app/views/admin/sites/_side_bar.html.erb | 4 +- .../_plugin_summary.html.erb | 2 +- .../admin/users_new_interface/edit.html.erb | 2 +- .../admin/users_new_interface/index.html.erb | 2 +- .../index_summary.html.erb | 2 +- .../index_thumbnail.html.erb | 2 +- .../admin/users_new_interface/new.html.erb | 2 +- .../plugin_dashbroad.html.erb | 2 +- .../admin/users_new_interface/show.html.erb | 2 +- .../users_new_interface/temp_edit.html.erb | 2 +- app/views/layouts/_guest_orbit_menu.erb | 4 +- app/views/layouts/_member_orbit_menu.erb | 6 +- app/views/layouts/_orbit_bar.html.erb | 165 ++-- app/views/layouts/_side_bar.html.erb | 2 +- app/views/layouts/_side_bar_content.html.erb | 10 +- .../back_end/bulletins/_form.html.erb | 14 +- .../back_end/archive_files/_form.html.erb | 4 +- .../back_end/writing_books/_form.html.erb | 4 +- .../plugin/writing_books/_form.html.erb | 4 +- .../back_end/diplomas/_form.html.erb | 2 +- .../plugin/diplomas/_form.html.erb | 2 +- .../back_end/experiences/_form.html.erb | 2 +- .../plugin/experiences/_form.html.erb | 2 +- .../back_end/honors/_form.html.erb | 2 +- .../plugin/honors/_form.html.erb | 2 +- .../back_end/writing_journals/_form.html.erb | 6 +- .../plugin/writing_journals/_form.html.erb | 6 +- .../personal_lab/back_end/labs/_form.html.erb | 2 +- .../personal_lab/plugin/labs/_form.html.erb | 2 +- .../back_end/writing_patents/_form.html.erb | 2 +- .../plugin/writing_patents/_form.html.erb | 2 +- .../back_end/writing_patents/_form.html.erb | 2 +- .../plugin/writing_patents/_form.html.erb | 2 +- .../back_end/projects/_form.html.erb | 2 +- .../plugin/projects/_form.html.erb | 2 +- .../back_end/researchs/_form.html.erb | 2 +- .../plugin/researchs/_form.html.erb | 2 +- .../back_end/writing_seminars/_form.html.erb | 6 +- .../plugin/writing_seminars/_form.html.erb | 6 +- .../back_end/web_links/_form.html.erb | 4 +- 53 files changed, 1308 insertions(+), 982 deletions(-) create mode 100644 app/assets/fonts/entypo.eot create mode 100644 app/assets/fonts/entypo.svg create mode 100644 app/assets/fonts/entypo.ttf create mode 100644 app/assets/fonts/entypo.woff create mode 100644 app/assets/stylesheets/icons.css.erb diff --git a/app/assets/fonts/entypo.eot b/app/assets/fonts/entypo.eot new file mode 100644 index 0000000000000000000000000000000000000000..56d104e980200dca00b46718db1d7b59f9976fed GIT binary patch literal 24636 zcmeHvd3YSvm2aJ@uBxu8?ylaeyVZ--Ep=PFtZu30P2RvZAbG%WZ}I)0$B`Sl9x#`nS_wcj7c&~CNo)PLf+SYzf;|9 zOV}RxzW2}k@mi{?yWexqx%X`MR?WOo5FQ&41QJAnf1)5ub^s|45r_Jg+7ChSmeaYc zu$#jl^DSRqO*{8=2}8n|uwB?GY!Nmi*&{4R%7}1_Fe(fRTagkLRw8Aquua&25@SNG zFjZ(1I*{}Jx=7;AP4p)TTNW%>(Z6=i>tlj2up6oIg_m5qK-@;(5rh@>h<7ery5!2y zzu5jD@^3`EcHxy*&D+>>>raq>jUX7xk}KOgm;Cf6DKxqR`777nG`Pk8)d-#Ie6b9i&+&~8CcmkPq19qWe%*S+_HL;s1q|AVW0JrW!~3{{{4 z;`Qrq+P)+GOwD_U119ml(amcI4+oz{|MG8QhJGKaCgj*3te?slp;7voPXRrN( zAT0a_((c%@dE55l;S#bgTO{uEpG~5Wz|S*h&iqUy zdK*z^g@kAd|8Jl9nb0N*XU>!ak?fYZO7*wjPWdcI;IWgq;c}@tl_Dex^pGG6PU@p! zq%EvG{~$z#dne8luIa5*;>XhAOV^4vx>T&>UjayrAqc@y)mLi}U6IKO&vdA>GPE)H|QdvP&b zoX$@dF7ChDZ#B&A596x##ZOT1Z2w%2VVQ-C!+dey$8c@K#kk`(O=KU)PYRC;hlGQ|3E`NqUpOp$MmQ-vAsiJR5e^9Zgr|gY;fU~&LytfC=%Irrj_p7EnUhZ(edNHtr^b&wcKq2- zKl3yI`0w@W1<(JY@M~I1_tTr=3h^27JCak%N;gQ`rE%$$^sb}HvD|T|-Oev-x^|;>Tsx&Z^p*N?V}bFotIf6F^#ga>y~2IM{cBIq)9cyj z`76(l%{FtXxyw9ZzU2*i=Xl4w`@OGtPy4*SIlgOrJA4oMUh=)=_xdmQKj8m7&#evDe{fT zk0QT}dZRtj8>6Gq9npKE4@F;${y18SrDOAAYht@&M`JI?-inoK;x%nGD{Hpb+*fld z?u~cHN8|r1(UMr2xIgi=#2=H%jN%M^h(K z&!@hSdO7u4>dn*-Qg5YBr+$ewaMCSZCmY3OtHL>q*v%Up#cO}a@ky) zsHh$(DO%xPG1n$}MAbt{vceW^d&(lo<+c(*mq}zgLKNM9cv_T#O28HHYhI6{yCkBZ zT})RzUd@jTB}kE;!~3HYa&RqIFk!h-4837ipZOEy%IiWxXcmM(Caok0He69ZQI#}M zk_BABaK5gydlnV4jVm{5acy{&p_rw`ra~`2)TPJt?aQ2^&q?D>pGeoX_b5vDR3#Le zenlJ*Hc68FWvAEc{Iac0i9(MMp>NPh@D%1+ZIm&k*OeqNVlS#`PoBj$93JjvK|S6OaPT~n7uBC?fIskt_5&03hM z|7oLJskb^C|066)K6l=FPBBd7IX&c%iQIpzNthvrGxq!ntE;Kb}kZorq zwuxzvm?9;Vlz`1p5MU%QO}h$odwabllj>AS7A09EXRu{fYtTumMyjfc4lC2K{R4wygK%G`SVuI?r3RyI-=+hD*w?r;c5((=!ercRQ85c>O z(CpO5vWb91iXxLFiVh;hTR*kltB{P5V#P^;M3zLuW74{*G^qO`u1Hu7_#B!~b`tqK zfyTx_N^G- zqI`dUzW>A0OLq)+xzmsJ^%d|By6qMMv>QCtgINoGK?wz|a96LRtDvN_1zvd6wegD0 zKH1k=7hgMl%9QC-=(#JpHjGC|wJw=^Rl4gVjnk$~nP#u$Jm%vYn2+Gk4QVykUC0%~ z;DQBE9?hy*Fih~K&Z32AnA2%@EVPGcPbgN~+B&DT5~$@wL|gZHm#Qvxske1@Yimm3 zz?@bjDXCVZVE$)>7Wygehm|Ez77Htv#c~5bE_N5dlCG;wByg**f9a)qx;2rgsYxX2 zZm4_u>AD**?sIjqR^MNno9ZeQyYrpgSiaK&8YVY5`^b@SFqDebba&UpQla4cIv9HR z;b6MXAB?uQM}z+PfHirhM8ClL(9oNeSFM~{%;ia#b^N}LPkf@|zT@=OXNNwwapUKP z0Nc5;0hG?<)QqJTt(-~4oK>`nszpCPKC4pmy8hbBR>kA1E~}-`v*jj2bFP~qj*N&i zuCv!=S#AYbW?+J`;9^k~DuObzD4i+Gvb@!0#|zp0gP~+H6zks&&LQr3CzO1lKZcZ( zN0Xt$;bdQ5GTcTe@CJtthhWaJTy6o~_X~bR>Z%=_ZIH%Gr?|%T;I2?1~qAyA$qR#;UsAqY4YYWWMK-I;1>No9z#b&XdR+BI#q>zn7L&|R1@ z4b2Uy`m6euE-|IaF?-I&1RtHI&k8ZjS=R02g+P-B{MuSs(6%0I0VYm6E2f z4qAivTt7IkxnWQZs!@08z?|zYox5tqV76u6;Pv+mh6YtN;(7GCxeHd!g?=w%ZT%VR z)C3(FOAse4Mwnw@(ud`Y^}I#*S$Tv? z_h){Tp;oV0*WX_!dgDQ3l@ZvzJAg;%9tgxNG!W7zLeKjR8Z)j=&zYW1E@9$pa$}}UZ_94U{6{4yBRCnGVzAlZSB_= z+8SqOlBwo}Qym(ux0F~m?GL4Wr3e2x(+5H=xg6$4?u5(Oz5_3Zb zptvIur|eN3x+FQB4*G#2J6y`rc_Ei}QVD5~`8-c+uwPy7x`DsZm3UhB`DIb@Xi(^4 zw(lYUh3`2*73zf+fxVQ>MzKd>vB!Htz&vj z?eO`ds^XHB9=Glm4Ndhq+;TxOyrRPwH5F*5R&oBNIjLs)e3l*2Grxw7^#fsvaHVjy zE$5i;!A@uc#NE6o5N6ruAv}571Gn}hV9pn2^M#)%i*OjzqL(KK<^~&Qm>(JFtGMW` z7ENo>9qt(zU&^5wm%_@6Dt;Ag$kk|gNhF_Zip!x_igml(j<`cNW*aeQO!X_45_Pq} zUau5;vw+|6oEFh?ng_N@z_rq?D$S~Dw785$!zl-3!`fGWun zAg5$_I+c~df*K5iq(GQzg|E_54Z$YMA8rboUaWBOlfV00SH0_R-EMyDJ)r5qL?W1W z9nHBHdjq7x45`<=Pc>w{TGZ{bG)4DlHTot0dxxs-fD&+&>{cJ zdc6q^%@S&$2e1y(47)QANT7MxJwhX4?}t?9t5adc6mjM9UCm-BkLV1sD`S;YGF63; z3cSKfr>L-d_wK%7{>A$GqWOFj@l&Vr`8>Ubd|s#_8V)5(BQ$(Ewi^Y8W3kw9EFXi{ zGI-|q^e&wOpYTFXriCVkNnk$^>v{tLvzXMQL-ap(*7_H8sSlfq$5@dd{5HZ4S5T z(cnOBM`?bAK92y-xGR3OWLVWMn zN~e)9dg!|kTodc#_8j0Ze*%Ab{p{R;%@WG6X3FNq`#}6WlHRbh^Pv6s8>gsU%^RW)hthvs=K$kb(NCKYb+wiuNoYxzF1cs zF6j~N+P0s-ozOzNTiE6Blh^WF>ucgYqDxg>Vo$s#z5Q2CS$0~ouk^Qt_H1*ruNjUq zrgaNweK%H53Y&yNQ9;k(8~&_P?1TcvB!*4Qvy@HEZusw@Y(WFDINqw z@0k0GwO*g!pU#Z#hnpeaHd@qzob*p31u-8N;Q8;uLoN^HY)))*j#P@faS_ zr@LG?e6%)}nAz`gf&aR7k9jk8#;^-T?}7Kt7A}WG*vwm-@E>K;nOt|iJ6FI`F{uDu zDy^h5=@8_88f+sRf&gG00Jb|jT#L~D6%4@EmDDzo?PaX1*@OW=^~bP42{>=220?(X zj$WFYM-vD`C>AYx&o!q-lIR*ZYA}nhlrb0t7mK0A`$e;jwH=?sO#M!jI?+Ju6tP8eyRVHAotJq05BW2%~*hCzgt1 zh8v=V;4vb1L5tKiC+j82YR&}`HzWax4_>3X!6{@KdadPA$7JsYO{#DCYlp-A;GLq_ z+;}DHW4GLQAJ0MfMbG>b?7Y7N&w=+1f|M12uvPZ;K~T6cW|2B89wuMUH6{^hXTEDf#?9he8ff zb~ZQz!GI`grV$WB*9K*eLRqQFI;5ZwXA z2;cO8;(_Yu1h0{urigXwkX_!mUv;Z~PsAgot!7nha%&!++mlutGMtB9i%K_3ZYOz( zoJPjpOJ0EWx(oL8`7*8udvUffzbxw}s+(#)l(E3(au<18(28W3v;54#ols;g=_jQ{ zW(Y>#OSpeGpYO{L=kxsqzU4>w2Kh#@GZ6Rv9E**N(5{gAE?2>=kdGD$eUopynOSmD zKZ`zleV27-6CAlLtIw0t4tVxhqc=*Bv=#NZ^nlkeyhg#r4|=#X49Wcl z<>@(=+IA^O@D9T?4AaX+^wMwMZNwo38-|Jxlp$bUp|gxVmIm;wWV-6u(^=LS92L$m-#Vq)5s=ShC)2m+Zv14AZs@`>%x(`oM z@4OQ?q|!Fi+_%p(RlmAz8*URr`xYU=@*=+iuPe#&pAyEo7Ub^bi#*ie2>e7*-GvPWulHMv?P z4-7+9BqK}|?n-KNJ^I$BUj;U%sJ9~CE@4C%OCM`voneOOF&%`#$@m#z%+0>M*jWe#I@X5RbJ5RIM zojp&>?*7U)R+q2N!s?Q3q0;c~@o{=iQ)kUmwbEXDMVWbC0Pt6# z&A*|;Xw!%Gg+QniN^u@6ObvC05y&bPo|wP(&^0sHzp!!T&RIti9qDvOh7QfV=Fqhp zUs%6$*3Olij~q&KD$kof(DX&n)NTeAWn z;be0%+58Lp{wl?_(m}h%ZkHI!aRpkdfQ_-=QSmh)&B*{B>j`kyvf-3%Se`= zS=DLlal!)7Vi)GE3)BGSpvQ6F1Q8DPz1(LG+Y?~=Z@fc-S(+c*bNJLdx_L%8b$CyZ zcI_TE%o23;-KH@-KD-;_!x*vNz6^O3we=uqmXJGP8NLRXP%yCAqt*)n+5r6sg0qca zd$MsOvBm^Bko~dv!-`W=cdO(ImP&Ux1M=On%XiuxkTOX(?-@yqN7p0Syr|y0h2Z^l^yCAodNg+pIRn6oTcwM{fcBTeW!p9Kf*qz zgT2@R2*i95&aY+&Zst;j?CHxlYG3Ix^}jQAE?;%$D4o2-#^?(@p%L@=wBWFLoI~_lwA4>8=;q;A=^gBjn&Ys6 zbQ3~A#{TP{0hfjLpkuIoUa`~JB+u8OdHUULk|9q6KUfO>(BRT;bx11xi%iG93hCMn z?PwQNs-bK!_j6fb8XwMH7 z3cBghQ<2#ISR|!;g5Rs@s2Qd}eNS(1Pkq4kx)E^bs;WBz#_L9~ttOeQX#;ZDcIE!; zy&DQRp(<< z=iKhURBo7FIE~g$pGMmVB)>5{eEKvCss8@bcpl=ZV$wdn4sA1juEHM3GqexuPAIa2 z3l4PbJ+b15m7u1-?=x0wvX#(g9dyQHiu3b&fLc8}T?xO#we)vx)thjYe(tqT;gIk# zVHNGf**8eIh8zT2Wp9kP#;`xeE)Cn~O`8`i(EL&Dl1sFxUz-qqaM?0W>F?*<31P-< z)rZb{l0fgX{Z-Z=D%(A_b>Gf<%w&Dx9roo-YzI)D*N$yF!m9+zO>S?vFpU5Awa7`s zAJnFlQ*%XF9z_fLXA@x@T)en-$&yz4KDpu2OVxDbqW!ARTou4F&g#`*k#z0n@4WN# zcdjgz=&$j-^YgO{ytjI;JaBccJmyPzXrW5IuM|c`3L}-O+c-P|BoVat*XV%o9{1J& zOqb7fV7>1t{SoIGbkcEu&mD8)-(F*za6aS!&WAAFt7{A^+TQMCd17y6as5rGv5A|g z2Hu+#vy(3hzonkVWCi}uRItZ5tc>E?b2x1UZ1M!V#0gK0Ek&ys$;s?ovnZRQbdGq> zp<+^OEq!&Ye}2C7r>5e)^Yf%Oja5v64%Nub+K)NGZtmmwosV8Y#Nl&Eoovy_`L-AONsmp=^cMPN! zg-zhC2f$l1@MQ`DTnp^^V4T`61w>%Ufu9#)p#u+aE~qSm^I>?%X#dso^68?qBA(8~ zJgVLnbcxP+DoJuM(%=j|3MP~YN0K;#@v+(A%$1SW`mi^sIg^&BQ&C+y^!|jAOus{k zwj{;d?8ChiaT4Ma5LIs+#9+3o?2id~s0cX*$RJnPSHiuN9937 zuR~mhMcx4_@zyguZy)uKj{0lznZ{5sXf@P_aK@msz7qL`tOi{(7Ib&FP^Z7G*=c!Z zU54V+Ram`xmM_#)6RoMqcozA>O)(CcZK_t|DK0$TqNmK6vsSNWz7mC=_XhMlFCuka zss&>Vf3{|y1#A50O`Aw4y+CUqrD*Q7>sL=3U6c&(ukQ6PpwFQ$coO_>JT6=M4S(nM z+rP9Zxv&4`oBJPsk&9wp3PRbYHduQp!UTrB%*-fPw8b&&^;Yq&!nLak5R&u0%C`p) z`_z5I!{g!Pq9orCMmuGgSvR%yD#-0RHm4b@w!*?BVpWA+80;^-GCDkr(+gz@ixDgw z9i6uN`f16~{&2D~4n^n_ZiGH^BgSE2-L%8PUkDg0+>+0xJF~cVQhs&{i^!oL4>LYQ_T`RxCTXF~|^9G{H8N%E-y z#3*EODiQeLa|vhV7NjnZoS;IQrC02SZYNRigq-)vBd-j3{OnlevpS??+lssofjJ1n z8huRmGM9HCSik$52OoU>j9^a&`_B4E&qMfqtL@FIen)UD@ZLhP^)cq3jN3kAtkOUIom6wI2{W3xAuKIJqFCvDXXqoMRh`@;EGHPDp5 zK>l5l={&T&CfDEJ=5qeK(`Db@X&JlblGJjH4=>6<*Zeci>*TOoBA{cNRxnI#g_xCM z)@RF0s>1fb`eYoI{Ig3d6*QM^zUEzQj+et{gxz#wihD}s^M!nVLhz08e!I`A{%@k! z^YDqo1~^xc&9Mi(yw}{74<<#Lh}wP?he;K)rp0Pb3X8H?zXk^DRCzDwR_`M!FO z>Q1Gd8V*y}*TSeS^P;Lx46}xjJmFAgW!nH+H(S{zw%~tR>hNk565;lG%%_VX(&|LN z@q7%&sbg1Sn^gr5Y<0U&yYa)VM|1g2ziG_lz8t`Bz&@*EV0WcUxD@^Iq0G9p2S*<3 zz*ju%jmYM#9JrY3aqu<(e8+Y6?jt**aE>PE;f)N`Vig!-=kqRADZMEN@=09LZNx^fZI8rZZ)2mE zbO=tD)BkNPime2f_H936nxuam&7-$`qrJ{&!Jqsp{K*DZ8}_UMXInmUVkbSwS>GU+G(>a%AiS18(4!((c;1hxbSg~?gg@B-Z zpf&^Zcp_0r*t25IijB+iSFe)A;CP7SRafViX}Waoq^+dWb9dZx&u5DBrD@Zo`9(Rb zFVQ?6ZHbPfHs|_GaGwqTWoOV2V6)&z?|y}ImuSiPG45B6Hw20i+d(4$@Ui6fL|P145|AVjKFw`~gL4YS|;#3tCZ zCcPaq_0lV718LFHFj@mVz+smVrf<_Tpqm>#!}cm#U2y(jZ@5T}y^G=gF8%U~{FAd6 z(=(k{!PZZU&`-~$ zL1Tdfabh`#?K|)8^jEy+{z|;+!KQDwwB&Q-3)Xkm2YplWEv?>AeI7w6=*x%wm2`hN z?+Z8($=8QU`$LIPDDlxCpRw+=?=NsdXlO=lPt77V<&V|Z$NVXEQB6t-gFzva@ceux8Dj6LzNeuAaVr{d7zhteuJb+Gww8UryN9hQ67Zw6R^TK6z)G zySPw$ex4YL6?VdI zU1+C+hur$^t+e3O34jKJT7<{^IgITZjE%jsa129HKs&iCPUk>#2><=rI@PXebaKn((UI^K z&|p4j0Nba5x4$Gg@^0zSVZ^^nTzg!-726+2v8E=7dkLCW^w>WQd z1zI+#x%28b@G@e!;Q+8lK`&2kVDkm)Pk{ftu|LRg+Ta85^T0pn7h=tqbq5JdK)@=y z^<8@VOQkP;$>trxQ#eNy#Xbe^nhJh4TL5PF`=0CC>+0I;9&W3vYpdI5hbe;K74vC&_&h{J+r&)T!d?QF?=c&a+TtWgroB~#L{VZj@?#eR#CaAD zyoOW;<9z;K=GIEkU)k%npfbV*ar8Xy?0P>3XHG0cV5gmgrI`k&fVZKWodZ=fa~44h z0q;+H9BK$hg?E=;nD5K(V}qZtZ|Scs zH@DOq#zL51@D<^z+9uVr2h%`P*eKQRCZ4N<9EQys%3FNohIgiHWd`0x(^h8kh3cNU zeZm#{q%{5x&eP%dIoT75hw{OA*ucRdy!1118^3gz3g2nF9e=ly?AM)*(Kvq5MyKv` zNPbPLhh4K8dew30RRwtUE`bLCZ~5>+t2XQsK}WK8%4~vH(k7`|XA88!F>9+!<=2ap zw7;_LPs31eRzM^2hgY0aXgHy)B6(X|D(O>fJ?0HtkNLG7-Wc@xJ_tMgUPLjak0eGMlh=*9^) ze~5hw_=$oW`1u3Apy1(sry}q3vykI=@dRvW;?FkpIAx(%b%%{;d(py)n5GUoS@X-8 z;aW)zB(Ve=!!!4In+p*)Ig&0j(9x(+bL_J-!c%02mh`JqDm$^Ytg0^ z&HmeFT|PpgWL#4sR?6+u@wSHOPt;|97UnWGzRr~2^^0Qn@a#NI5j>c18sxH5(RX#m zmM3E(w52)TKtks{|CS){t6@?S{C%a8lfOI8+?a_Y>h1 z>D5=mCcy49?4Nhgo`i4$XC4~hn}xRrYuM_-RiCwYow49}OL{uxAzQnAl=ETB%vUtw zmhDnQ#czz@N6E1J)mPnN?}Ln`y?xgzfFSTN|8RT7K5Ey{*mz|8Zc(!~`3!Ci?m} z^aZ6Vfb%DLHJ9h3f9#~=W3l#0=bvA-OPoD$!)4-iv*)f}JvZN2-?QlI+i$1O@3?Bi zhB?=){^;stw)OVsZ)ZA9Xb=2HWd)n}Sg^$5(^kF+_OQ#dg+752$| z@ECrr74IX$+0bV9U<(v}sdb`7hHcJb9_QI$sKIptmQ?6g^I@xvD4CvOMW=+ly7L;n zcYUvZjZ^oAr$ntO>4tV3qit_k7!&6>+Ml1(?)SIzZR;E{*4Tb?w7$mlXEQx1ryh*Q zgSs=-lgawcn)+zk5jBG~H9<4#IM%(iea2C*5AGk+@AB0z?T*`P(`E?3xBbKM&;K8< zA_LF=p7=QPHu%lkEDd~ACkiD*N)@18`1(bef&^*psf0yvl7>oH!XCmcm9PV4o~ndp z_|0Ccgz=#V;cf6Aa%f0#q3v>5M2n$HSP~Y1E0xn7D08?HmSG=1UkUT+obOHkRMF7b z_MKZc<9j=kzgM(;XylgB!L0%s<*yZOx1U?_MWZ3176ir(YiGw)6q?wUPzCHFpG{gh zv~}Bt&11Dw+n~f=up+R;^7+5|eHNK`7anJOaUveaf)!E;CN1IZ4|0(k?i~}yJbmP+ z00k*Te0V=XQHoIw4g)8kx1^{RKHoG>xYSbv_C#|ug&L^|9@rLY1=ne(4w_1xIQdwh zF6yQt^*}$Lj+5mxq5AaVC7L-jm*&B)yo4^L1+1tTA*U~Dwj#k4bcLNR3jWkGWXf3UyAsU8{U_EW1jkJkI=_VQjm)}A+(^lF> z+vyfqaktS9obT90x6?=H4!V=>f--kEeU$E@dub2d#~x_<7`{I8aoS6tpaa!dz)JZ z)(>nR7#kQ0tQ*)cuytT)VC}$mgg9-Rdegul!lT9(q>mv7Vc!;{aeOP{qe$6etV4Kk z;5JS(Zo>Vhfi)b~Mv%UF;Ff_czQKWY_?wVtjv{9*YE}BuwgHYo?mw~~bvd+sVExW5 z>xae$hS9|6!07oIYsxjk+fbHkOfLWaOPz6JMf?zMp9CL6M`f$iGONZpLq zhBU?sml!f|*{!b~Snu9~UN!&|C^4{A2V+}0q^yko(SeGHp(}kCJ$>}piFQ^5QjEkSli?+cn*OiTw|L(^=$NlI5tuY zZlfAz(sEl&CMKO7LjOGVfIyImTiFDZ@=yVr!KCAom=YSdJIW!XmXVH;7;7e!SaY_- z+6k#^&!(<9R}bi{K*b{qaUa9zbE^%*xhYUdH+g~%RYr{^sxri*16$m;*$|J}Fpju+ zG2Mc8hLJa9@S@(1_|}0jkhL;-&7pxEfPxntlYhv@)8>IKevquXkjhlu9-NfKm_hj@ zPcbI)She&|qfbQzSWjg{9=WiT*4A1+8k0Wod@G!#rp!wrFStwQ3gt?9m4^xXk zql4SW@U1`2<8S_fha=cD$Q0mqTkj`(qWOOSJq;5&jA;^PQEf)ZTs6No>K+=Ga73fo2pk^3mhJRy7% ze)=bn{|M6X<(huW__uYhAp<9>-KZK6><7oRa;V;oMhYkX_ z6H`wBrYGL$>?aXqp4v`5Iw2 z@X0Ol*pJ|T1lqL|-$Q&9urk~X<9(>}4Dt`yFfB)))zP;gT!wT%EJz7oQ9KBo27&9D zfaeC7Y9B%EhwRZEw#UJkJ#1s)2w-rc_G7qslspEe_W{88hzL7Wa1;--kR1dN6I7qT@DZhH(~PVlK~+^6WT?QS(gT z7I_@Y>==JOe{m-Afez&GzMhXuA0u;8?Z;dhuzO_6v*(v*@&VL60^gy89*?1}SIA=3@=O}Vs_nrV;Mu}7E^5 zW~Yv$=jv+lPDrb+m8CW;$3goO=;a_d0@H^_#&g|`{&~jpTDuqHJZ|HJXF@F~G#$2O z3T++)RLn7|>sYhfIEsE5yQ}Rnlvm?jXoG2T9BGU{<_1aNX|=Embozquec|6>&o{xw zUIV*#oW4Ncp;$lL@!W%Om|Q&f5Iz!p@!Uso=H%kJPvYA(7tehhbJ{%tW?mWmEmr1hj|z8QH17oX z#>MJAiElSwEbkEZ=r5M{IPiY4yo2x`xIrI=?h(ND8GGLI{5pYIdOPO8IlS}CIbo0B z1m@(0bDk7#yGYJKxUVi;=TTw(MRFd;7osm*=a6tSxY;?_${dk#d&>pW4?*@_sQlx? z9T!YLA?&$e`jhyu_=Vazh*S-;`;g6Ts;lA#th@3WJB)Ab9t5|W$bl|FjAbwLorCa_ zpW{1D@ct=Sd3{*@>oJo)0@s~BVFo=3Dm{Uf*JN`*mQ=^F&RAa6BZX;q5V^)v4j4s`0ro;f$?AA{VK(b7Rk#mB&- z(%_MckhUJH@IK7%Cos;Jgs%zT5q=_^7JiFSXsF9l{)7z!k94gK(E;?uQi!?x<>;fz ziw^>aZ8qL`wR8D#;Zuk+hvcsqv$k>BB+Okp39EYn-?%-C_d|}I!|AmX;O;}1N5isk z9m3o;!;-FqPom%Hzym{f1iLI{?cxc%H^uVw8N^v*X#rQRLx^Py%dT@6ZbUneBZVn4 z17R+~oKv^MthMYz$Ytw)(}BMtFv5DpVO%`IL&Ad?(>^@8&4a>?cC0E-XCYSB5}0y- zaW=+KNN6dJc`*XKb9v^MkD>pA_`4n@S<0{!8;A7e@jU?;na&Y-3Y~z9C3t_rWIulVosft!Ob>qo<=P$;fLNU+jM66d>z7v?cNR`SHgIXLz_JTnEjX~&)Q!t eA*ASlIC-8Pj12q7C4_H&d^0v2%HN!?KL0OcZ#1m{ literal 0 HcmV?d00001 diff --git a/app/assets/fonts/entypo.svg b/app/assets/fonts/entypo.svg new file mode 100644 index 00000000..01ff749d --- /dev/null +++ b/app/assets/fonts/entypo.svg @@ -0,0 +1,580 @@ + + + + +This is a custom SVG font generated by IcoMoon. +1 + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + \ No newline at end of file diff --git a/app/assets/fonts/entypo.ttf b/app/assets/fonts/entypo.ttf new file mode 100644 index 0000000000000000000000000000000000000000..0a114e0a62ef8cfa651c9128c3b5735ff6e3c152 GIT binary patch literal 24476 zcmeHvd3YSvm2aJ@uBxu8?ylaeyVZ-jTk5uUS>000o4kQ-!19JMU<^XGC3z9Hu?-j# zKm<&PLjr?sYzu4%_83EeOad>%WZ}I)0$B`Sl9x#`nS_wcj7c&~CNo)PLf+SYzf;|9 zOV|v2-}`6&cr8`c-S4^Q+>N!Yev!OH%1b6(pb2m@nCjW4|D;sxRk z`i>y1Y(Ttg;j*QdkN(BZ2ata~;`IwJzhd5|-kX1d{Hp}PP?lca(Y5raKS>D!?Lz*l zbvF!d^Z)o{UJx9}r-w%dcWgUz20aNYp~3a<{@~Dm zBJcm;>e+ws>*77)?FNUdBx0c4weq3Ci%X1sGaIHS~%jpac=S|GN0`ZUG z;{4jh<@wrNy*SMM?#0D$aXLR;xVZmXzqK&8Ka8u^7e7J4v;A{9hGiBm4)eu*AH%f+ z7vqlGG;wj6+Qsc~zqRWy!VD8%96pJw2J`0-{tB*Iel1+1TP>a2+JfsfT($9YS`Dr> z)vy~+hPjr`@L@*VFfyD`Wc|PVc>eMguda><5#tpX7aY8%1(&^|90+BEHevT~v6c`P zgfwD<)!2xzOAx%zRa+1Q#V+xf@PzP)a7Z{PoDhx)`-Q{8XM~f&aG!A3vA!m*;{XI`r5Rj~qI9;@JMf zpE>#X(T5N0dvg59qsO25^wUoPfd5{XAoXVIbm}*$Kc>p{O1-B(T%WAZ)wkEr z#1tzCNqU){6&gWsC7;W;i;C)zlA;yv6Z7qoM^rtOBr9yujwdaWTy857beTk^BSg{t zho?m;s03UAzvlHQx=SJo+QoFmrqJgM3I)S6sH>;jSIo)zc4{H3(A6`WdWt``WM#GDlV{2J_aP`- z4n?8FEZM7mHZ*SM(5s4LPM8Y6Hk}E}imGIQk{sX)h70Mgo>^4PHLcpD#kJwphGLeNm?t+9_*Fv&CEw zsqK`XO<^lZDWVta(@y7I`(wNJ-$hdS7tLzOt+}kI;U=P8^>nCoC9wu>2$m?9;Vlz`1p5MU%QO}mS9TStQ>lj>AS7A09EXRviv zThK|WMyjfc4lCQa;kW;`#n+HB>Y|D)nx@krF&Z>t$zU)UGumXIX^OHEtus;$zFefI zCt|OUGRFS0kQc6j@Wa?aB^x5DL6pk?5rBtmc+?~%s8;5`fI6dk!~~_gi#ar}=+hD* zw?r;c5((=!esP>Y85c>O(CpO5a*2RMiXxLFiVh;h+c34ktB{P5V#P^;M2(?$Y6V-W8tt-3f+Znn$`JU zGPU)cC6kjJk%msH*J!+BQK7%T(EnlSr8|bZ-I+%hFD~LAbQ==_GzOmP!K{V8poD@} zxVz8MT~so;A}>7Z-gMbkpX_T($Jb4tGG+P{diJvJjpGqgt&8Sfk?H=5!j1h4v8b3B~H$+UB%X1NEGU zXxl#TGS#In^R~@yYfC8{nA3(NCDn!$%>S&=NK(lzyo z1a1utFTU77HzyKxb%{j!y7W^|rLV)d&(+0ReQ#}Us=HX~DRgmTg)R$dnB3s(BS*r) zP%2i}(^D5qg@W(tVCbQTf|;~G813kY2L1B^Yw}E)eu4F&p*O3bT6wjUFOV?n_`RK< z_(bQu$LXuj41I3Xrq2xlwsU0zD4or#SxYThd6Piuj>lJD zQct00DoupuTr)!)84+h(W3S7K+zPPFzyxE##iA-y0%c}VCR>qZ1*_YR7jyduL&;<) z);|W$A?|rQlzhHFhLn>>lcB@ml`{9f;q=>xfOKZFZdBj*E%@cAdQz# zag!s%BY4ohWa#-&vcKPcRMrdEf06DK0^lA(pfX9Uu&QK35O6Hi3KvMbBh}iP$`a+$ zO_vutG;Ox)o9Cs_otQC=Esd#$D;6(XYD$x1_MD9gJ~~aG6=ImPm_HCYjGQ2g^lq#~ z#u0Ax^EaGf=NqbQw1}F%e#1sxX))v=D7yp@%>PLmcFS6zYR&5cWuU1$>^fI`azM5>2&CZSaNN z5Iw5attDiqTTzYsRxVp>sNryQq-}MlC_2Fk)7`6@BvFrKx@InDb~jtelu8;Y-Mt-N z3d)kMn`+C}wSlBB>92PulEd}>fFtmdPc{_UABiPg0o~9Y{L3YBcq}H~FuY^z+XdtGS}+{V1?>iU z-;{&P{Y{V+)!7M95@3FZ=35zY8Q!^guI+%!{uYHF-0KE}1 zp~uRBkSPkY9tH~AxiyuO6-p%x%vzYPsL%)f!7iF#7EFXBHAfD~VP>{Q#XgVXahbXo zWrG39;dF)^qA4nJT6C(0N{)veN~0`)U(=(8wuqb}1s#Eu+lktms3>z)iBxwc)8b5c zdLmA#yg=6@nsIxMI$f@)BvYCk4o9%1Lnne3BRgfop_9iaMO9T2U81ZuNKRFU(&Nz; zSdj9+>smg2fcpf=*|hN7AHZ+-OX=8m3Hf>XNI?>?=Q|=@F-KTdPff{mg0l z4ZQ_>rju8+tzE0kQNRK{(CHaT`0v1SP!(G(w}2pC0KjFKej~bEvKYV8qnXQ1_ir>w z8FhtzwY=_NWeBlAKNl{Xmf&E@j!gkV`wMgtSL}o~Ja}uP%3b z;BRy#p3!}NSyVh46uKB_4dO!Kdsa||2BB48FD0{4tkhCA(?JX;@UQS=qQN6pm6o2G z@=^~qVouvwpDBA~^Et&O;?c*_h{kb$cGG2-ZMtlp35H@&fYprIyUh+$Q##CK2%^uv zl~1G0Wt(;x9>ZXtqbM|?jJjNoaPI24KEvs3_64+r6j2;LB@ywfVo0>eVS1V`%A`Hfsm?@G!X;|) zDGB#KeQJ#^*Jau}r^nPzpFgT9E?Mby>u%A|RG-5w7bU|hI($)6fp%(@=3ktb>ZZ?U z*%3YSYuH#n5S9v;3s>55j`<$!gmysO!;1o8mVF+=lczm!Yfl2^d|@_U_=&OzhcPYs zc!FSVuyKa@k%hjBi{5P2v{v2Wo`LbD9GY=4th}h=SHXr{O@^05^0}tC9D21_kIU_d zJ9J~V5p%{=zhWs-S1auGYOyzp_#Mw{5iPHIV5Rpv zBp-^{#Y9C6xG#5UYJ;j6Q*<)g3@3!6&gC!>eYx($9*|Iw;1Nr}`+Q9B!6ydowZvwL z*ad#t=O{rnvo!&zk}Lu8N|vWnMHwur!7xY)gsE2e3N6(bY_|O2=Ah}t3Ku{5yT5fc zxc=7d=EvRxnjTCff*IG*ynBf^Kq}0T2F?3qW6rBZ-7ZU0bdOf2U-Z9ssp>8$0k^1X ze^&HGe1T$E4gS5X%kDXsdd#4Ymn`JK;xoUb7wG$}Ut?-vIfeodNtHDT$AfM6v2o>N zEt_2t^z>=#Qd5=OmwdvZI2^k6J;e%|-*lOp>ZUcfxYCy4E&pC~wp}TSvQAN5zR7Uu z%4StYs^(It9-aps^1q_jo6yiKp$>Wg>mV(#I}3mWnupyZG!piHNNHbMg%wl6l`nL+ zh@k?aGsNz!RY}R#6hbQU3M-$Y;@H^O;$i;97B7w#3Q@#QohlRx^eXasp@wKUlq`?X z@afnX3Jk|$vEf)D2CrrC%L*(J4{x$=b>8MaoxOm*Uj^KJe*Fi z+lB9cWdlXJIjQ`IR7evD6f#(P3CqCKwgHQE;A!A!9x0fVy85Id zJXuwEpgoBdLkZ|CdWbr%yK;JKFxWc%%5#DF^RJt~U4oz>Ig(Cyrejgs8f$D$c~Z@d zv0C6C=B$}BXHC1qZF)2~(CRJ4!J;2qnCEZo(6Q^c>)&46u39PymNm6z9cDJrl zaz&j*3BctXoPlk<&mKt)?2KU-irxk9n=M=liLh1JZhI-R znQXqN(33A>shCuRE|pO-*-Qv>KLfTA4nY904glMo9j+y4{|W|R>q=_7$o4YU)oj85 zp!#E2pah&ZQ-dHtcV{0>Euaa6Arwm%z3ZCODoJz|95tB5m&+Iof=fhEf*#|zjIvfL zT^BDlMAUGq)geI~iH^>uQ!Yca+jg8+- zH+Gv48)3BX>cUcS%y2`r5IjcYE^3i58NS&ElroRK6cZ+_wpQsU-ZmB!Or_T@EmyGAV^sO2wN3j9|YxWiB-tF zM{L(wrhpyM*Y#^k(=O3u&pn#tBtyxYTCrhPeWrWrWvtODF2$pmvH=xPlDvf$jYg); z*Nw|d_3>BT-hu)Lo#eSfU8ktpPzVlUQB)lU{xn%f<*;7g1GpT54q5+4$a>bB)(D%C zR$bkYFYOg84UTKJ(egRlSSuv~rmYKmM3A-wmAWhGfL|TUCm3eY2NSltXnXVW?T&08 z1^QqTV@zBW`80g$!3R%0*wCsfaI%V`=}CK4s4%iq^m04;*@-TpF<&sC_5XSfnY!sHPZ-)p{s+kN1>e5?D93L-Lo_F12q5Oso{mwl%U6G z_xQ3b;vGn+^@P-zbE36GFFQ!HoG;C&+W-5 z4jInF?nUJrCAX8jL{1}X?*_-;20^w@_GI7%mk0i+n4L@D1{fVrL-k`#BaH8KK=F^Bt~& zTd@!=78g&xjWM(2q<$8C_WG{q&L%i=MOL3Dr5*6>vB*(s?g%!z@kg(hA!)1Xaru6) zVR((Aiy!n*c^H!Wbt=%aEVb=Ykl>w$X&9!Li|FOwywik33N{Q?A1Fh>xd&>A?y=z5p|D{u6=0L~N}FuI79hc|BTALzF=1xSmo zfBBa#Ko5y=1t72eJL~s+qAAXVWWQ z@3CWEud3d8r@9YMQt!9}H>A=I)7-buG*!R4V+U>%L;INb_hV1vdDvD8JPoj}(L)%X zw^9*nu3ds#UR9D(2$G5Al5MPJA-L2RGr3&m!B0jq>5wRenl4^_aZ^YVL+MPUB+cD6 z7hH{3yzG(MT+OaF$pgbs70Cz_g}YO!QcG(o8SiR~#oD^!$x>@eDOK*mPw~K6;!}eT zHiR5O)$25M4`5sjpUw;LN1FDou+V7R{W;Zk%wp>ZECGIib8!_4lWl3)8%y+zcVPDb zT&~916?cDi8>`!wwy?VBhvgNDMyIs$DlDB9W~e+oHauPC#?3jqEqwD~u57;XB{z7Pm?K`AbPg{h&gFakNH$`cD$AG&JhhUYh}x^>o( zL}w<`nWaNBuR3(~rsp@@I_uU|TaO&da4OH6#h~d6psC#=OrHE?%|4D@g zk~T1WpJ(;@SNMA^dbYnmWR`EUGxnGvd%*%!9)DGLQb7rH)KhVcXTsHIz7Y;c-W#Y` zx`k9_qcLr}_|lnN&v3FOnQZxmeSd}GTKS;jSn5!w7G%ZYCSjgKDZk}0ob#HK$=d&d z<&yr%7T#oR`eh`?&z$PC^*CVxXt5jf)&*(+bI{|sZ-NMi`d;ZXkL?LC{ny_n!7R@Y z?m2wwZQVR0oI1QGNV~^|4YLd#eatk5$A`x-K8z9T?aPoyQCkm!W(m0ymf>rF2?Ya- zJ!*XrppDRvAUN9zwkI1m5^GG5GyZI>3tKcgeRuo44%r`zKcqM{bxb8!uw1^~8IbRi zC2v69c?W|3OM$-xT>mAl*(K_JpEj~V5*x$KPjM7PM)1_bj zkzp>|e8<%*c2{Uwvy+^dCGh@2&W3yim_$%U&ju1yV2nQB8yYc>PYVv4$2ml=M$7&5 zylx(jmEXqhs5uTBNH-w_WbD8G8E{!>4>|_h=T$qMP4Yq-&C~C0kqmhn_`x#phenrn zvqMtpUt~J=RY=z!Xh*xDQe|}4KJ0{jT&e75FUELZz`9G=)?^ki@%o0!$MOwifuhEA z;a%iB%LwGrfsVpZv8bCKJr#-Vk3~|tC-}X(&bna=H1ziM^)>`tuNeV{uBy5tV7z7o z+v}3ax^^IkZCCEk-W#G?Z@@&QNP^xDV-Y%Jz zwKZi4*q&(2`dE}yR!~L!Fb~U?@+^Tu6noK=K>I!>3QPkm~O*j~5`GswVAI>(Msj z=L+nBJWczs?t~IMxZps?-V-a1SP5$8`#xihCR+(@)enWO zA6UMeQ~LWkcS4wPTl1l_o+QwFY=4zCi0XEaZQXaU9y3{Ac$WTk>W_T>JAQ% z07(Sx{Z%?3yvw~c0@Iaq9a!&s%74Up2Ay==-*v~___x>CW}FW>fb$_t_u3l6inh1= zSf1EhSzLcJYHa2vYJqnr#q8wwh2K)oVzLVVXR6p^99Bnh^*Nlj0XBJpUE+kN#+ITr zjO1l@u33~zQ6^8k=TJ2%wwJ%UrGI{*{HNy9J@X5sHgBn#0?B*&w=|dkRG3?82BV(e zzXe?Y-o22~H$z4nyw?fs3N68W!Tx}hfn)(DR484AE-+5;^pXYVN?){YN+ME64VM-p zDV%<2Ha#6fu_i048O{cuw{;D@l?cUhbrEYySy+8{BA{FOsVhF7_IX0B%-1IRQhcEI zInPs<-T8(4w!5ymoBnqy63CF1PkH>Z?1*a4_$+V44FTznXtCb2Zl>O(X(c_8@^`wI z_zcBkC8jP92HY`_S`;>ew;lj*&BB){2yiX1=Yw%-yA%+CB?o?9f`tw|z`3BZ2rh)- zA*20Q&MRa}(#m)y8}q1od(b61=cy#g!APSs^az+xA{CpQVMl$m@CEAk|Z?_NkPQ*!wPe9bXaS(&K?utJqKf zVMkIoqU=V%#(E7;zacX~kl}p4ShhSk15$$V=1M4%=2U~&_2?Bd;rgb!_*tC&K4^LZ zLDTdMCDQi2H#dzqP0zGRDI*6Xd( zoyDtH7a=6)eU)$bBlf9#hlj_*$wf)NA&hn^FtcuI>s653X*Q=BtG2?zBw|&Co*(Ql zzdSlTjMEDh35yXd937pu=Gtk=(Ef0;Iu1oxEL;zLgfi@LtY`tn+*nLEO+99qrg4GLH7{^3-(ET!vZMBGpBXaoKVpVW{0~=R zh1$bT9I3>B7l$wmT8`WFsEmWpCPIb7atA&N6T*P+g6UsrZ)$39`h4Y%tM)KMJ z7tV$d;5j}Sf0N`>BZyJV;Z!2}9U(K(K!IHxE4U+8M!~4ECM%k=_UK`&RoKHT{m@Sm3>d zQrn}Cn>6WTu`{D6P-!yiuGy4ckA&9)K*2F_{Ytp|Z# z#oDW9dBI2ERGIhMpbO8SQi{8pe|B zlac&8?7mCh7Wv)=kLpfkoEi>OH`K$ZuJEFoPYkn$kv!p0W_8;DS~pwSCbr;zS?=^| z6cXX~deo@343g>8o9^S}6Emnacc0TV? zmGT>s(?jbQpTeW;Qg1cOD@{?C6cVK@$#t||(tHwS;oVdu{VsC54!K-J4fizA^?VXn zbQ`e|Y}+Go*xS_PB^`p(<@A4Bi()ImrG49vm?r5TNAu{d&}6UkS@0*n0)MiB)rLK* zKxLP7HYphx@&J*IQJ$|A{;RSR1ab}RCq-ArDRa@{i!v0S{rxY6^;mmM4|mSn{>1iq zi|d=~TRK!lFX~E(LjS2+p^&BiXZ9-*(~LOtJ~`3coREF_shP|)r_(u?Wr^)id=|cm z2H+Ed1X!{1ScQO~a-cQ~^LQdrN!YV;?aEEd3sQ|tS{9(9&M?Pqc-RIOmLr#|7B;;4`8$4Nbi5$-3~ysXU_$?d;0FX zZ^zuJ_^$Yjk{fdAJj+OBk3=R1J)U=1>7Md0Y|aKR?1-t4>*gu;9iQ{gT9-1r9gB{D z=^cY^b1ig4K7YvWh?2#ggZZBbu31y@KCD^8{s()l?0_EqN$Alo`@|7SvnJ`|3J@Z- zsyjA^3WnKlequ9hT9e+2nR@wUw1Ko}c^It$9^kNB2-COeY0%A$o?&~HtZq1eus2+y zroJU`f0uuGS>cJ)C%T3Fh1@K=0;15g&sGyh8eM&A|sVO3u%3<)=YdKCvi9{MLj;N}qCp&&2!1ST+I zwv#8{U9|Pn67L47F8BftL<$X|^8QdF6iR$F$Y-p(?ECYa5E`0M-&?mxP5EOD4KaU8 zT~yawKjU04E2s3%#BJx#Zd|+ejtM){d)G|guwgo;3)arWeQmT?vo9y? zYeV16PTJV6RG+-F&0SomJwMM9z*wWc$|HizvcaR6y(`XjY2F3lBfL$uiAHD@a~1-w zhfmR}cWwrr_X@YdZe3`ngNNMw&ds#o)a3M1A;0&Xn|GfAT$9=fgi7O;W{`do_Wx4Y z|2`CgqX#mDPa^T&1{|u?xI@!67|w_m2wZ=CAfQEPbeA*Y+@d>43xw9M5Bs&M|7f#t zAMD8vA&7Q7Vz>Qb3HA3M&1I0(C5QY_8TTNOEEn1T18nadkkgSeYnf(u`9b? zM)FlaJ>b|Y909Gew(V4@RcWIK0*|hmT!Rvm+Iw$(`xGVj4R@;z$;rvpk4>&U3HB1S zYWVSZ=L*~mj=fhV6@D-147UWhzjEPm-^c^}ri_1#_^&gvsmtHOxW6Uqru*=zxrJ@0w0??|UR(hs$# z)9vYfc9Mx#nlPW8s5IT$cca5eWuUVPd$VVa>k1WcOx?^>5 zAe3SGGyDm!0T!HvCdsb$IH(^ee1 z(P_@SzWKoFJ!8wPhL+X_!&nIO3%(*;HQS_i_Fx)l3LB-?-NbWski)QfLj{YE-0;qn zt<1pNXxhq5p;+59w@uiB`IA(2isq%V}lJQr!{b?BL%}QuQ{_x6k3JoVzR3vX(Yc+j}t;f7>>oLE!!<&LW zzsI@GS<|d0>Oksm(!I9&b=X#`9>fEq;#(etNtlYTLeybM|4#zO=7; z9suWe6LDx^P`9thez)uw1z|SA>1qBc9JC%5!pM@O1izi@16Mwd$$0-ZFsyl2% zJ4zN##58x($+}<84A)C)Ac-Z|6rQ=q+ft0U$&qxKfzBp{T4JA_5uPGDw4`5^QsEhu z{$%J^Y{$oKn(UVlCF7bBu~KfIj<+>Le4QZ(_yq0 zrmso7?oe?+-%o^7q*q=En*h7duz%i0dlJG4oOx)3Zx-GjtYNDgS3}O;b;g3@E$Qi0 zfNbsdQQn6wGhfMsTee#bmA)~8A0@-?S6*?4y{8;FXe7y}yk~@C$9{P1Sdw#3G-aDU z*_3HI<&~E%mA&rri#~O^C@xohw5WX7JW9@eb1tRk{?BSLd)~CeL;gMbJmwAXS>hAp zrM#^s;WaUASn)wS)#5jO^QBL=P+Ma`TvV7T&S?E4ZKW9#6LxltwKY1swZg1t`r4ca z{^P)mi3v7dP4x9|=nG0!0OwB%YQBJ%dTl2iAB(k5I{*BN-Qw(l>n;(mnLT&Snz@Ch zhTcV2-gX;(e%BQnH_o|g%}3WHb8WXhcN^1bLVMspDl6E$$ATpepSBvdou!oeunqEA z(%*wlr*L9uD(sW_;4%DKE8a(hv!UJY!4@d|QtL#E4BMQg0?xC+P=o6PEUDO|7Q$9L zQ8F{dicSf8b>~%j--bT@DyQxZPl;MnGL0QLM%&T2Fec7%bUZhw!|(6l+x9tPtf}M1 zXhWUp&t-d4PCXcp2X$wvH=FaDbq&#sBWecg>Vjs}aja)q$Bd(1AKX8t-{osq))Tka zrp*w5Z~KSipZ`BzLk6DzJ@IkoE%2MSSQ_{!Eed5s%2l9U`1(bKf&^*psfI;xlE!LS z!XCm+)vyC)o~(vt_|0CehVh{X;VtkVa%f0#q3udoM2n$nSP~Y1D^=1RD08?PmSG=1 zR}J&&obSP_u?64yxK-GO^Si>(mYuh5!*d0aM*zsEFvy2`R}78ZG&+dD0uar=rlrw$1{!3d!5?&7O@mAoXaDI~P_BZQrqR>z4Yd?TB2kBCy2r`M>&o z7MXY#9%p-TA|A(r6;cT%E#d7Ca*-SE9TUeqedMPA1t~;)ct1i>icuX711F%jq^KS~ z-waN;G*BbY)<#LO-95ljSp^`t;!?nmII= z=E1JKh%TlDw2&6jC3GoWhA$8G(-PR^%V;^Rpv!3`T|ukpN?5X2(`veg*1#us9SzX+ zG)QY{9j&J!8itQx18t;Dw3$Zf2HFBHzm0CB?X-h-(oL}9ZlPT`-?5u+qmR(-bO+rD zW$rHeDBVr>&>p&%J<#+qe0}8Ow3j|X_rsU>DS7}Wy)VNASuQNXPS;|1#ungRpjE;$ zXmU%1)$meZDqIEY@DgFAummTEuYvdMYT;V!ZEhRbFtBxC%fL`z{lLb7?E^yt>jrir z#A!R!8wLgu9yPWheG76B_H9EN$G0Osij-}}dV~iDZs9cJ2HbBLSj%B;1nFA`ZW`F; z8yr}VzX^HfD00@JR<$o}2jCdw{v#VumqR-THr%>x!_bz2VKgy1FnWH*+DeV^4wU5@ zlgqz1xmw!I(hz)tN(q;5rPLmFd+OAHyf?AF%}Y;bQwFB^dg zlo;5q@+bkA%2?Pkuvr~OxzT|UncE!EcH+Jj@f|WlH)8N;83VwTkDnumpT)im3Jly5 zV2m-YZo<8WnYDI1n?Vi6{K&u%aJLai83L9W+FRwJfm;DFQ-g83+2o!^k-r5k0WZF_ zHq=9?!#NlaC^Lw^Zp7{padk85-8!&CJ4*wzlEP3hj@P@nwXL8w<7GQAaueWS zh(JM=N4_33gd2)lL2CMFfq`8%+2nSw8^#egFQ%K&&M@+Z3|`bb5#K(r1!S#G zUUO(*7ogxp$K)Tf@w9bdn;#^rEu;!ncLpaVF=kLc$y1Dpyroup6Qjf?$MC?d+ARoi z7dy4JwGieCld1yBkaja-m@5NYwHq+t%39;0kJ_o*-P{oaxN+cS%-EfX@s!yBLNfY! zimbAmVCeZErSfYo0JDJ2_R`x11&G%^#4k88 z3+&B^gJ0TgrXsB>9JI=E&dL$68-rKBy*vkQRd3p|aY1J%G+q7?-av-IfshB6?FdAi zzj5#}Jns`ej=a6XLkRDK=8x}Wp;Sc^<~lk&OsxWq4({B7Z~bu|f8&pHXmsPqAkq-B zFNd8=NFn`2SP@;q_o*JbY#$vJd&KL-Z%KQlUph89UQ|yzN1dP2?$y4f*Xt|wUHUkB z2?Mg}&RV}9z>i|*fu(d+(%I6ap9Bj(?5>E$JDkN>9>aGD z4+_(eei%=@?J;FTrA#}taL)(uEe~F$sbP%|mn0f**J&wN++B|^p5tLYhHVz_f zpN-=_z;PTcJdXThc1k^ZI*4x_vfal<(;Vo|R|#XlC%437KZ5%aXxC1B5AhMe%5XD` z_oB|z$Uk7iv;uwBM&F8X1=9VnASHZ7@gQ&-1g>WSp6g($eFU{1v`2T?9tUIgu#JTy zfWe8{kK*D{@)(%j2LR*4Hr$Wl`;C5^9!CHVkMf9cKjH`RJc)Yz9Y-f%8%OO=BXwL@ zi=KE4CxyEZI)E}qgpUF)4v8q4gymqM^%F=L$G0H4_X;$#?NL>*CBrUlhE=f=?F|82 z+}~k*AMzmR!L((Hj@y_S##w-gxjaKEv*RR2%`<^pRz*Yg4{lL*V{9FN8Dk)5rN5KV{T1>b7m@5Nzk4$;?{PIjbfVxNE zJCxAlG1T=6IjmZqNuyY`y;uW0TbO3t8n4qM*dx~L)N%A&TP@xRX|=Vo%!cJSXnz8| z90W&T`tZnjuDj7c&v;&I_h6jIZJh8-s0W3n!?sMJ&4Yl7IYw0CwI#_aGc5@1J`Jo|pH}eFSGt z-aq#V_(9%3_c6?A_XL=EW$?FHnXf$}-0{A7C%`w}uig{*cJuq?9l{>{`{g|byuV-G zLHG~cpbta$Fkt(PJ@0vboxm);4Rhce-g)Mnut#tLbMnGDPYAcXPtHMjjV@g05n;po zb@#dTPBl{I!4-`YJ0Za0wwU4$6RUgkRo;Uz!Ecbwq;Q?T+DWA$&qO!^3XZTf^6 z^aQB%I96V>%>h|b9mhIjdDVavrrkm09!JYh0e>u49zlMM->q%5=(jJ+?;<5a#Kc22O$+71&_*rM=nCz2CTw+F~1+jIA0XLCVWTu ziEvu@Ek>cCE=&0nHViz{bv8r?&=*S~=I)oGj~XvN2pqQCc;nU1<;R6jAlKG_@dytI4`59D@Z>fR3fJ4QnmnC_SVc=<%KgRJ7(*eUr8wrr2=LD3nO{DN{tx2s zT9jld!%}P<(wE2gIACNtN8l-R0xp*5wRsW*j68=Jf=Zb-lsSkwbxsDi*tmHLwYY>I zdarELndS2}2p_h4JAhmX<2ep(_5@(|W0pK)f3<{=q66aOd3rE1>>rm9zWMR3*l?Kq H?fL%;*5D*8 literal 0 HcmV?d00001 diff --git a/app/assets/fonts/entypo.woff b/app/assets/fonts/entypo.woff new file mode 100644 index 0000000000000000000000000000000000000000..57cf5b47c2a3e902a0c1b30ae29ea331ae9771f1 GIT binary patch literal 32144 zcmZUaV{m3ow}zkC$s`lowrv{|+qSJqCbn(cwr$(C`Q<%-PgmW$>%MyJy{gua?yBzI zu5zNHa)18F0RY}H064%;Tm1_F`mg{0O-NJ}9{>QN2LK$a0f35hhftSkb~YNFxpm8WQ_A^V{9TM2P&s=lN?p z*^_HxDwXMNimMt=%FP3Mysm@m>5PrNqPfPZ56OAYw|4OR3^99j{`7kH;TiS`B{jVu zUe?rMXxr1F`Gw>5^Kv!l#ySnN^p|&TBv1f!c|sR;x08Wsack>tiO$FSX(zd zX_!`aR}QNrJgyzpfUP>LLa&vrW3F{IG&VRlKh{RpO*YZg&@@)oTGw3G;nZa|Z8q^V ziL7tAv2)LP8r_dvEA;IXMRVeAyE)yRT;FC66Gyk=j&u9muI$z>?`Md$6c?LPPQzxk z@`Sr3Jysm4Z`Y@qYEIKn`)BR*6uVbFbDz3TUzyXHRU2O0WE*Uob?Zgnj_oPWmM1RL zS0C!UcU*a{cfz>)9&Pq}mOYbR*Kf^kg~IEKX^4S~xrpJ4$%ui7If`M4$qoM+Mjv(> zMu9dPm|7dgVo1Dd&qrU5;k#DZ+a@JaoB~K^!vr9zX zwqw3F9MeX(sr=m%Z;b@@+sT+FDm_pPg6EIKvuw#cQ3KQZ*#?4y?Z?@86_=w8&mwT+ zAz$XAM=HwDw-AJj3d|BvCkdS(*P`I_OqbT?R4&e{1lfXhu_#-M;mOWSl$K4E*7H-D zWmFdk#F{Kz9nsl6kmxk-`?<!h z2J30#ld)j6_G%qk>fBV`ZC3RwHNMAuM^&Y;!|Oc$7Ed#|N8Y6H`#!rCnu?|DRr&=L zjLr&+jnU}!R=j3k&802vCE7zO6Z{PDbLYNk-m?9JNS49B^-(4PV9rf=tGoju>P^trF)LI zfnIy3;K$;Yccv$A(B*O^mA3wiD=-Z70tA4(+?@9o=cMp6RmXEwkvHZy8}_lNr;q=# zXl=R1;Uup)dwe2;SlxuXGf94s`#8x@Dz!ZM^vti?UQZat$WTBr&IS3 z52x|lmeZqqj_bG2_pRp-ALARE-R;I0eox%`c|>nx4ZkaYF+@ER9Dw}y7T3Ppe{nB- zzy`(NCfL`i9;VpWd_K0=nGc7{y4q&YBKN|NmpN~9-!&~KpFOxNU;hg_!z#Uvz)*1lRGP-7EUE8FwUa?(y zxdd-n+&I2w{i%Oh`vq#>GQDPtQv1H#o+U}CjIU?jFnaW=jbDqX5^=Sd8 zt{NSU>IyTep|Hli3b*sdL_HD>i6yLYV~O+{b=LIB{v-4H^$x2oe0guHE>=GM-?*&4 znv|M;G>tToG?g@&G@Uf^%t+ax!Fwdr6by z9amw?>%9ujX4Y6QFrGJ_q-1q%T)A6pNV?Qq)v$({MD@sM}4)#1-Aa zRkn6wY`7G0kMsTdw?*QZv^aBPBXyd?O0tqdte7=r)?g?DBiY0xW7SFJP#w(_FZ0tV zYCD6_@oc)1sru-n42F`{t<&xtArMVj>%ru!}MNhL7U>*cpLNmA}!4+6QPT%!A zj)JUkq}j3%d5};}^3`dDlwxkOQQ|enREVN(Vgvs$-?F?=e7?#bd2oJMQo%fo$e6Z3 znSf&YohaGe)RLmVoK!s4`<;a1dZ`=#uyediaq7h>Gl2=sS%`7@_h^G{%mjxY1YBC~ z*k@9O7jE?@=M-~{do~(oW3#~hWl=5>tPrvs%ls{CIHGrFoZ-2i=xdHGw``O)*Sa|;a6wC(UTy=t^V#3kb=BwCKF21hu>p8=e!+7Gf;ckGHePKsv*b6r zTxWIeYa7GD>f8F)KgA%KL)dm}i&RCJntGbDtm|m`9sL;J+y#wnSvVS(!keA>vv21S zBP0ful@f=so=dQ;=2G&-JHIHbUR9)04=O^n#@kO0`A_E*kvCQ`c}ON-f(n}pofdoB zJdLLQW^oI4%~N2$ZmK2{o-Rr+4xm$Jvzu?B;eI}qs4$F3?K{i$Wi^#wX|M43?u1a{ zKi}F#F@l^aTluGkCG96k%|6XL>$o*tpojUsq>f6mOzW8|mo$oxap=;(8sN1TdflHB zFgNOh1!Loenv-qfRw_k&gsV_{_i;ptO_HoHV2_KtpT-}WM7iNbBROUYU2RBJ$U7nF zVA~yJZfw=6*;kW!7|)6hE9jgM8NBdQyj^G04tg+vx96MWm=;*Tv|sBq{7MRB`VMM^ z*NQYeIkkYP(!_V*gZM2Donb**iP}908+W%Tzi`)g8Z%o6a20STMPw@#*}XN!q=%l8L=K(}R-XCK zGWWNJq@zSP?V&cAv{~SwI>^LTg${lO+Agzu6>y0Sh^=TYej+M@LjG2TsNNivXhHw7q=|nZH#H3) z!OdpHqnFOY>Y-nDkYadQ-Bz67OdB9p2vg8?)r%Xx8px@;QVU_p@UtH2M`=ysgjsv2 zG0bez)eP171gVnTO}1OEssUe8l@;W|Nb(MYe|r*xP?AUGeZh3tzmj@wd35=My;B(?`5AXobLAj1ytd$IHjln{Yj#2wSRoMl;QPR>+=Sn_?0KQbHcnsl-z3<3% z09`{gkamrr6ic!~_mq7_1b-oe_qKa~5O2Ez_EKYmpr;tFsuLp`ec>s4u5+YVLq{V#w8m#cBpkvgo3#&F zHf&%hi$Xr+0|F%8XjsQ_o7#ciV{ZnrK*$rV(hQn1>Ef#cJ`zcRA&DnJ@V_3OZ|5x! z1P&;5hL^q~qvk{w7|oa-Q;HNty_Y95j*9(h5AKctcDQGtes~BsQ4??CW5=ia$^jzJ zr#|8&=)N6dO7R1aG(vwJlg?WpJk`LDsTni{4!S(UHgx=z#I?SPE`W9>L6T>m=FZSn zv3SkG>(KxWq`*45h7LIB_iQR3l^0(J%2T+X{>mz?5rFj*pKXpuTUr*;)t3nnVhCEp zgywo9p*Rp2gLUu$*Gqkb3NS>I1NWZfYhdhabe*}vL>lw~2qv3~WLxESf3HedYmJAU8K_V^%NL}M@Og5Q|vvXlc% zK)3}61ATbU@vq(Qca~_`kzKDaK%@Z+4qf;RX@(n@2uVUO zW9e{2YTmQhiv@GD*$*?1{d;QPkH3Gkx`N<=7rnw$r~Qr$?PblAPqQJ^AY zUL9PQDJDR@LfWc}068}~wcuFENO&_3A~w;S6|tI;OXTWskm+ON`No31zjUya#BB69 zM1$Y+0^7t)>-q*E;(?5eE>uE$l>Z5Jx$4-V@_2~VMU`@+(r zVZ>$WlegH2yiJ|Ux9=eDok>sE@BO(j14`+Ps>d2e3y({Y#?(Po3w*w=Uj{@MZzD6M zYqs1mbtE@&vK1iGlFI2|$~sF;5eyAgo>Iyqby8FQ&FxzK>zD|&k7n+%-{PwUi(bNE zYdPzsU!wGT*x_0oiQZQw>2CK$%&QZ0XyGfY&CULAu#plDBpwn(Xu3{bwW)98tUL+b z1>Z2i&!qgB5Y-QE{K^b?5Z+h_R?O|{7erm$2;G>p&FS+e75o5t$EQ9A}tT4q2$f5hTM1hW?iUFhiY3e`j@ zJjO$fT08ycvcsqh%r5FmYjC0;(wr)X;f0#^iW29FPVNTPdqSmzERW8B>qqZW-m z&YM1ixUEKaoxYcYICp2()_1?NIqF&4QH8Qx8_Xu~{H`UWzl%o&GS`H}=zMk$DdnEa z1Ycwcg?hw6cGd>waAo{$VTGjkRN>9p*>NG;b*m#*Mm=xEgYDjIX2B!8`|aw&4L|(( z#@HuX5gNPkWmB1jYvQA23!{3eBgJ3k>#z0)Y?<(G$Ui=G%>O=yj;N%iY<1X?9wM`EHX*Ue?q6Z)yrKCe7Dum=Rf9qX4{ecra}^@ zO3bKxFu$Jo2V+@KbOUmZAJi1nW$N}X1k!QiR z&a0J2OIBD1x~H_9oEt({@H1|4*WhKFM+?~!A)m{;x@?baDNBo0)<pym24)b{V3O^92j=8#~ZZFYq%43i0S9TiLEw1|7-LfyQObX zi@m~eA6}n`p|9P39 zKFTjPO$|#FT_sPLkbp4ZoPSG`$qr)qw2Dz&8oY=dn(2dsnX*O$(U((_3AZ%G$j4+Z zoB}8dA3-`;Y1#2{Kc^|RF^Omvc8j_KM5?Cv8Q6}B2MM)52$BQMZBPpo3O`bZiK&Gi zkG0OWDm4FDx0z75)!+8J zkt86i76}$860S8-t!uIdjH!HM@Dtky6Hahdus+g0`?L2=^1BhPB0{eyEdl3 zOtDmk>ldVofmI;VOZ7QtD_vS_kdpKvLD*AxNLv~BGTSfyY4v(EpnuX#OyM3+WB4CD zhGLi>Q7r4G%8;Iu>BR_dk!vBQy7{Ef!?lO@L|aHR-V)PCSS?1fESJt6AHCP#bn&ei zH$d<4!2N>rxLdp@4OqFVSJL=&<>?zr#0Tf_H33W8+~Ne%f~VNM7yHeSBG>CMc2uyI z`nw$Ok%Y?07}adqp|KV@o0ezf**`c8SvTM{8`L}f+i~JBo|Przt+}JF5Ggn@bNE!? zH@}XHEq#jmQdZsltF-H-| zObx~kS?@Kx-SOr(H@`C9AV6Jri_eb+N09_}U&j7kr^)JVR9(0ha)DcarMI8aTD zmR(?;jS$YZgj`1l$j~oUXYJY`e(D*dt`v5v4~sZE!fmDhR5p7FpyA&+oNYMn^Hiwt)&+1WEXzDM!*%cAZF_IR)id21{cWm$L zs|qm0Vn^v=_cO>0S%*l?+7@+(fQ%mb7sL`Rk_B4ca3Zfm;c3%trb=$HU3|`Of7uQ6 zRWd${`BCNecT+=A?|&lEp)%FiVR?nnlVm6}Z;%w|#<{={wStjh~}XLY5p+DFJq zx?QwcZ}>_n%?vE@Ea>Rp%O5?*p4UNpoj;Cs+p_{zSDu+llcmLF92}s;U}^6nVJPnI zT&kmM0%45;i!tMe)xj2Q`MBA)%um%mDmPlbU0={n3bT%iUNQ{AnZ!3-iC2DHAE^_Y z=`|<`7jF`?sa%fNL!=bP(`Uum(CZB!C`%*-{%*}Y!xY$KeinBt zmQ9$V9AzpG^z6|vqWT3R+^p9tXnT|oCzl&%1!fQsnJ`~AL(WJ`FRr{S=N$b^`@VmL z?;n%5+q^HIStj6?>M0!w3T3%8s=ct!#LpG-g0h<6drQvK%uZ;FY84UVgt4H#!{lQt(X@8i(}Lx}-A_fT`6A<9D;MpVaQUAt9^zhgjx#+2K zC9*2%#sqCo(4-7VdjBvIV(bbW-DSCkQLFxRqmLK(P+zi3pM;>ZD=TvgbqdhpIRu^K zVuK)uVr%40^%ort=e{!wQ*hIV5Wc^t4v09}nk8LdiDGZ$YoTV-Zh@~dSRd77Q+{so`S%Kc%ndtFs z)=W`QNd381u~@U3l6?r-$0*n`4cPRObzTdHl>DPg>2yzM032WI^(UtF*6hY8l1YnL zz8}Lx75HZ{b|{AD=w3zH2Y!8^mgCi!4HXAj3ylYm`xMgWAVR=TKn))o zpSHihVu$rG)~4)Pc-XcV(avKu{@I}*5<~L20@W%tk6Moy%LN$M#cqm+wDUSNs+dH~L_Ha1v=j1QqHHzaW_&^q?o73>vG=sID_TMAQlKzF$?$@WQ8 zJW!q#TA(4wjsD#wmlB;Ff|P0;XP}0uLmJa1)+OT|XJOce6e`VQn6#qLHZH^&Vd`FU zYk4R_AvdEA6x^Dj?|LQ^1P7~AG;$S3z_wqZ-=m+OUo_7-{|T4H)r zXabdDo+VNG*&;=B>;#m4=VIm%ljy^hj$`6ikP&KQSe)R@p-j3vqrNfY8yqNJnk(ip3%;XC`vy?f zy8A(e0Xc^J2x&5zH|t=)L$A$GY%U4@ zt2xk4XmPgxgS#3KwbDhK`-u14&?_t4tOK{QQbB&p2gU1w(}IlxGYuok+CPvu_xo$v z$`aJBf^bF@;GKf^1~N+@pqG;8Raw1C89M)f@aHZG5l11vy%;O4OYoE%k{HY7uEQ0> zGKY*FcVK1ur|Fk^Y||t>DC0)TXYRVHr4$NOI3?aF9&1Q77|fu0@8&fDN#ZV299T+< zvRsN(nCnP zo<}|^R)hjBWyAo^|Q{ zfu|1!MQHB0O9N6_`iu!NzYrrk9jO85rzy~S51@vGA6*3Ir`d_p;l49wG_i+^5hO65 zQ)UcVcpWx95BgamCl|;g0mXVgkcug;Zo9YQg3mX!GMfCo@kvY8cEUF5H|Y4cj$Jv$ zEVC8(_d0I8s$TlDl`tihI8pdvy{o9F)^uiQ0xExX`L#11+Rk(6D7_>#)0#TBlg?Jg^29!c4qy$Oj&k;p-EeNYe(LTi- z(fJI`&a3RqP8abdFz@gR+WkzjmGobL{}y|R!(k6oaK9@r?EjeqEW&y4OxOcw8+8Vb zQ|ETj$!{!RcsLPid+jH1u*kljl9*&zG+`HuvslIB7DMjH^x0dG5DZfQn;b0fBdec9 z0i;bhI07whsMX|NGd%Kh`u;%$(xXzy(?c98X8fj4Ma4y9V)izYwX#wgu# zU6-A2hum;*l2{7UR=%3^X=Eu=)-bNrR5T-;+8Z=kyXpHH{v|9OM5mP@N$_`B?jKR4 zV*2+a9hul5l@!m!vDenmLVaS7gnHuvC%xhxB16k?N8E^Q;vBOv-M&G+zJFDcJk_|_ zSwaWvK6~l(+QF`l^%c~* z>7sdc`i1RtlwEGWL&j3@I8T04a9i+!qkbT2Wk`J#R=PBh_9=tpU~hsiNKGg-q_aaV ztH(NXFjcqZ^TR;la9?|AJ~RB%6~p$5bk#)$S$ zU-`qTWML`%3>6z%^q^e1_p$&3b{z*nZRp?M?D?y*g-TyefZted{{tIu|5@j zMe3^x5M#8mNTK5&me)_|%Qi9=0VsT#!V6Bw{-MxBDCBu!&XUvP-Z;cV(wR8eW5~kv z$4I&XtqCFqmKLG+s(&TwZZuPC;Ud%W_cG088#>fQIRaV7DrqDeAu5V!u*3+Ez_I=1 zFxFjXzW#HN0RywbbNCDA1?QEZZ}8^ElPdgSvdP>GMpVG6>r@u)R$tBq+?jZ0FDR^v zyqto{3M3Tplh}OU2a(C()0+Z%;3i5;LL-B|{-Bn};Mb zymv1PSC4FE9gwonu38uj6M-Df*8!Vy)GSC5jXzX!Wzdg|zQ+YiRWi>d^NV@J2sZmR z#uUT5O>7l`pULnH*Y5JGJ=;3?CvyTK%frN%?=13bFY+nKu(oi;skg;-iLJ2(j^->DccSh?B7a*{qBDCSV;0t*O2VSZ5yrO@{qSN7!7+=;s9M{!|uu z@P^>&NA24+Xr;XUZssw>7q)hwirqH;O`(%YXduX>@@w%8P@3^I?f_im6~X)h;2~~B z^J)^?RW{$v!V2@Mx}<_w^F;ch{k5BiX<-U}jTPS>6Vk@+zq^-XgXyPG9z{PzKjOjd z0Dzeb2#kYIS3*#rP#o&4D`H_ILu%uP|4faTJz9F|hcGprE!Q&-O2w1p^C&>*U!x8> z1adbytDpy2Q%3O3Vn{iVR#@Fc8q)dN{UZOeHTHhHrV6ga60x^Q!x z)Spc1?nTaOXDo%wb7N%rN6@)5%BR0dStdjoL`5;zI#Ei;CK3YCcgc+x-)2)?dKW-k zZ#`6fv=(XD;EKhMD(%<{KUySLpeu^=r#;S;nWV*ezXKIc5FzoUa6-uje>QF}OkUNK z0&MO^bv7299#+NsEVOsz7bn1rMeXlc%%zP1MY7rAGOgdK)v9YO4zczU1B0tlDQRB6 zJyDxl{yGIw37J-c7t$-93MRrQ8cY5w!!;8*gD+G)T=gl}+RSYzzJkyQqIB6RB^vw^ z*$ikf1}7Etl!-3-P?sb_U)0bpVka6zE@jL-VXTu@!&Jh!Ks0Vx$Fq)AiS_G_^lsA< zRqPMX*O06))-;;Yk==k62Q(#4t5D}HAYrS1J~{gFwo22XAi4v?R!~{m7{*G0XbZ6@ z#vH7IddW)jtZ~oq-k9LiVoo2zF#{_-O>!m&=f$)A)rm&S=`+Z8z;7g8^tBWp#3WBO#pCH?{%4Rd-AO+NE7lZ)8PV3-zapz7RVEXLm&g|> z+YRi}D$*s013X3|Jakcwt+5tBJpw@|w=?B+ezbdxc(psW$exNKp!7)EQVGgJ>)`^H z|EJfq`0SS|^I%75pNhCb=3Db+P%zoG8q(#F+p)8xN0|3RQOySGc8FfY`6H3W*;di# zkg4>D`RswrsGJi^r+3{W`cH&$)3^6(9e?Fc$42Vk!~)K5@++m`4iIO6B%h!XoP z4z5N?)p1(ls*=IBI~|1I^6P<+uM$iyBTpJe%K&LDruqZN2zd#~@{s=tIx@HB*roGH zPfz^!B6yLvNJs>#I9*1ISTgrZ%|1{zA>bsfda zHFPkd4%jxFG;?`X{VgO4W!EfIvx-_OakP-8o-v-F#3K^BUs>`y?SW8MeGcj1!>Y7p zh%ENKDZHON<^CUh{$6IvO0|#_*=k^~nsV6W?RxIQO3C6a?j*V-$i+unyjEi~%i0_B z*h4N}q{@9DI8qnim4K)D8y@~>=SA%@4Yvc(8tE%@npq?1QC3%HtQ;TE{X&Io}eS10&e?K@3a;x}cW}PHqTj^qVzjD0p&@Ke!RDpsr-F>hgtp1R;^2t ziZO-67I%2W3vn)f&i<=zZ^G-IN0Md2Ze9DI-+gx?Xt1PkZ823(rzhA#={tXIi7rPA zW#B-5gM|I5?1;#^WTc|2rk=2oqxBY=d|?}W^hwpo)YKB3R?e$+Gk%t1YT_C$^E$7j zb}TS+%GM~c@8z1vCvfF`1TSB80+u}(qoL(~PfW3cBk@ogZCzONO=`fm2FUWYRNwyd z(gxeume6^ZvpP9A;LcbWb`#+((zu5rIg3qP=G;8ZXU0ciA`qCIXV=S0jm?jh0@tn4 zP$)14-k$N}DX6PU0@T+Zpe&iM(BGU8P6odNu9OvoQ?UH!vwdSDyO$9mW-4B9DbDTz zN_h|Q-Yzne(A&a0Yo=rGew2+NR$vKEp=bSMBx@jO?M(bRo2c@G9` zedP?~BK8}iHt;RIXUesdP#iZ?%=gU>Sk-SDAGwL_k^#kTi)i{!PhcH)iYALx3~GO2#SEhaCNM@81y>D zNIoJ*Ji}juP#lKflNx$oosDbtXB#mFh9rdIu{j7HZ)7S`ACz5 zNTiZmba6cOcy1d|FWjIw9;d)agrSMWXe*y&E5CcO0F(5?Q6o>oNzlM1Aum1UF=z+n4Y4VFEiIo(@0EPf=GLore0sVww*IY6bT+p zWs_su#W_n`SC|6IGh2g={CTJ{L;!e2$cImaGXnFi-RKuGSasX)ZD%pp0a%MWwtfLH zc1)2IbIaGyPpE4tGP}(DJ;A*+5sUj6Wm zQxB@Y$1?^PvtgW`3Mg9synBTs;6(O!-TVgM(u1!>9CQb~aU51{{ zsUDh%w=6q;vCM(w?lGACjL4}QUw4Lz(j1)+JII-VV5K$_f{vzB;-0A})aALGSheB` zqO$=adr>nNaPbY6HM2@&DWoF_+vmLi^g#v66( z679o>YO!4NHb4l{{AdUiq+L9jFft5w>~O@4)=}|<>=m67jE4HcMcXEN`t3;wMc~o0 zM$f)#B!^&Wu9vb1da=GCihpekD#Es-H>P z_nkP=h2ZfXPv?o-eb<-%It89_Kgxl#9&j^PhSC(++%a3jUm(h^YlVtZeJL}Qu6Zt> z{b#1L0v?jFS0N#Vly_+q-+ba~eYrtuD}4WLa3^|jo%nHrv(Ku7x`kI~3A5DQIOSpawwXKgQFyR(7M#(g=ExEvq}vENmmDZ8 zM=ZbK{(z6M_&iFEuYOkI*hSr^r5k@QHBm{S!uWv<#4<(DF|q*wXF|Fy&La^zQu0q2 zwjPL-DW%ACO<0j%xoR%BW?PLIe~?Ane!|-w#Cy0p5IS_K?DWoqbjd38i`EWo+P)>C4#pa_r6OQ}>?}0{QwEVR}JT#;Ihn>@0aBX)b+B3m;)u#S*SwM9q zQAklm-uZa`aV{di!wZcmDeST^)7IeLBj-LelP2I~uY!PKQM9U~~9OPh!r-Sn!S6MI&=0F~y*ryv^OQ4oHBnogkACf`Tw?W<_iTG*T z0(A4;>w5R}snJutAnoZX{gi1pg3a9BjH$+cfi7Kp|5;-mY^W&~{NsXz){~fM0AIt{ zY$pxHFaX|2>sih+NI6x5yF7u4$DWLul#8{^sry&4OLXC9C6}1Fiq|3#Xb|U66LoOk zimIb%YIfCSWTbsl_!Mz9KPHTBnp`WnL~&?(!h_0f>EhhRFJ7V(^%JHm+_@lR&401` zco$4+Vdn91@+k~SF6WX~g$DSbnO&uz`VYCh(Cl6Uv2W!%lM&tV0o7k_=%PlUZHpbJ&pBA-U6oY&X5?1`dLpCueoB$IstV=EpfGJ9LDXZ+>qQ{a?lpn~~45k)}C17pos>t1V>YN;Pk zx9`vFiwhohtwH%SIu;If|7J;Zm=S8?C{XBvm+|BUvRzm02Q`5Y2pRBf!$jwN<;sQ1 z&gPS8g)FosV4|NWBuTFgwJya(1IjwN;3}<)2?n!0Yp{-^2U4`-{ZURo`_-UI$vyI; z2A5u2xy0DoRq;1O0H50A>VHfet~SKLYgf&L3}!i%u&aZx>fNX)pyzNoLTnzKPtcvC zIn%tfhL3dvTw6m*N8;tM_8fTURL34+7k2JTKviDAp|Hxz6Ag5Y#_W*G74%RjIxCoI zZ^|RI$KUR~rBf(B`8B#)*)K|#emS&NX+$B?lQFmP3B;1UGqWx(Pf|a`p2H)(o^S}$kW{ct zTH_o)zvuqRWE9`26*q;aPRshNGTO{X2{KKR(R4uxNI))H+4od}}$BEk;B8Xxx(k*}8}u>o{%aww>| z7NK$BtSIZ!QPBC8rQ=ybY00fW5Q+dh6RN*N1!Dk zBCSKx!@K}tom3L#1vm>~H`6R#CER*@Fi=bz)liUyD~Yo8%CmzZG+WkMgg$eX1tS@E z=jm!}j-l}x7B@dETbfbpE>i6V!bB>@^RGfH*P+dF#_XiyWI4lX4zhO7P6vf`PP5H? z5h)GhMScSzayg8!g~m{B!{3m-Cm^g2Y_KM|3gqN2z;dG@_BvxjYvzAR2P_4#KY+s7 zu8}>y10&SVF?IDe&*J*zgnEr}BoMRbw}&`9Z?P<+W8~b@fjN&4fq_2o|ZbLZYh2#^u<_rQ&p2~a71Biis){Nqr5fVituqvNIMu5jXlH!iMU=- z9o83QO*qq?lzm6F$bOb7+Ho#^8;Tx_0W&Hr7=Ow zoHZ!A5TFD@S0&Zek-tl9mrw{2?3ASHMP4QdEobI+K55N~nC8tTz91&?_)@KQoHQH% zCUgB7yI1M(oCB6IC6l{dgJ{XqBk~*JxdA7nukz=524}lWynT(RBiqB6>ep68Bt#-BmK3(|}=f{cyz2%*lc?U+GLX z8Ppky3j+=6SlU<~Fbw0=8Sj-X>Bz$w{s66L>>y?qdP^+a=9W?%)kxICt|yoj67|RY`XSa{w?tGicd2& zJ++YHW)TG67LYP6gxL8+G`Luq(6pAJo{ChWCQ0X>(z{euun$SG?1lDxkLoSH${&it z$+DLRAw$j|9kT*@n`4 z2p`6lqa5^x0kM{H2ESg-=>@$TM%0~(W*Y`@a~1CKs`+z2CAbEQT7a_{M&iQZ-w+#0 zRbsnvD=CE;)e(nr3&8ldI`vwb3xwPwIVH_U6Pt9Nx++0Y!4)Xu{+v=J%DkR0txL=X z{EJ)K0gGTs%7SItbUVD)Of=K!<^mJICPi-kLtxZUB!+pUYNNTpR1ak@8X)NVtr z{t6OZ`VN`L-sL`+1h-6`V~%K!_rZGB9BhyWYHV=JX>gIJ%O@F;>?_C3bgVqnpoi@F_6WA6tWQy4>jcUV_g?PwrO~5DlvsfJZi# zD&GqIJ(a`VMH>I{@z*^{wY??Ma3j#pV%t+iPY?a@FhQNh<#etvf*V`uaY(d=U_T~% zPUbELakIqc@{Y5eEq|7?P8OS7m3mjQSX<0ePHCgqreYAf7Y^=elG(JC-s-nWM9%#XyTdajM3+o|h zu2q2X-Jq)MA#gRKOe~D-WqyUDb>J@pOG&^ZKYr5 zjC(AUv{hPzIU>R#QJy~FoEy%xPacmy)~WqkbJaFsl0SQ?V!i~StuI~AJgj756&K4# zf;W&Avz*E5iclI5zD%U-U*d5$^%s|SY~-ssAb_T2`kSkA$D(n6J9tI_b}PrB}|!fJ+woN{On5QwAU(=bqZ5gaVpZRXx4M=SVB}N z5<;*jkG5>LUy8AhcrUbCFutf8h20eG{URT`DV_uId})mw0lO-K>@W=RsIsqi$q3PY zAsW{xn&>R!Q`*$a8NyjULE8+oXisgp{0?1eXVP9@)p+h7T3PpvXCzem`PT|43=R7# z*5BEF@wwov6{WQ#YfGU3{1h$6(;TVVqqM+7|GPPr$eQ=(z*(r_Fb^{jed@*3+s!mO z4;+J|N%T&*vl89t6CTVf!EgVP=n9dMJJv!b^4Ita*QdGxFRrLD$aUqZlRB5pJw$=n zOzOiSjGaYvQq-ct?LlX#cZAv5marE6U3i6?Bmx0S^gmlaG@D?3S(4B-Beu)RH?c-X zMkK{+1{u3e`J7BoI1r>QnIOh6$cy^#m^{|X`LOs>TL7vgtR`d6sjNCBTbyPsQ^p`X zW+%th9%i;IRj!1SZEF>_brBVc%pwDRw>Dm70Xs7lZLwv7IH>4sW58O470Z0!FDfWb z;u}%>8EEl*T z2IIa3be>RRiPM(m48E}!0n--3dP^HqpPD7}R^Y6)HQYfnGFD0!d_q2BmBQV@R)Iy&TdYfpN zs3kd;)k8booJNF65{7gF~LRTXC7GQD^8#jW%Q zUU$K#y{74C961|FuV$Ab2&#mw@J>-lC-{kOe5-s5F0)FXAZev;Re+k1`O+2j(-qQQ zyuJQOd#wxn&dWSfe{KTIsmM4-T%eMU2RRLqrQ#PP(<8*JfD&oVZ!sIAzH~1g%GZ?z z*fpZQQbm(1nC%%iv*l8t@2w;cdPD<5&Z4VB-31wgoqX0BFVA{Hx#~a)_$ShONWI)6 z>m$D0nGtULP&fKhMPQFsCXj3v2blH;jIedcXS>&Vd^ZlaYx zP%Uvb9dk(#7_E(t3$4N>FFO|E)DZ|&xKWXb9JVrBdq}56ep8`wg8UasxnZ=%GGS3E1N25Z0?mQOk7~xY}FKNQWzIa z|C{LV=kez$0B$`%jHTI&0LlZs7$nRO{2=9~!89q{&c49>cm)`>)gR%j8IT#*yzt0$ zkzD>V5D@cc>7PnxmLd)pC=b~%N1Rx)w^B?=1Y zztdr%@h<~BaATsWB1)jt(V9EK#Q5~6PkmM|hIk93;TK7I;YHx4+og7*OMCwY+r4Nv z4uP`F8{9zTo#{!Vv;}lE#3aYl+IzmwPK;w${etwj4M1%2?_G zltH(W0S62#Ve}HdbQ6rc1&F zN>$R~vjJrfCajazs$IFTS~Yt$XM$58{Ae%IZuvRTfyDqQ4QK?~ZVk23JLWqWp@cUH zy+1`)Mv8k$MNcErR6o!D&fGCGJ3q&=PTU8I`^PVFNEx9{PCV5dg)_U5c>0+kGSy2j z^WmU%y_Op-U+MEw-IsO~_u7x zz`;X}))anG&jT1_RX0V~Ak`G+_VWHBX)V9c=tCr~gm+0i^jO4DhK%D%){Ul64(S?* ztICyb4UI?RmJ#>8=~hMg$RVknSH;G_V)mz|Wmdktb_14P6ACfGS0YGOL_sdM@1U6pvx~ zfSxNfisU4zP?j-%Xf}#CjR8t!;^tB`BxqG=6a*ktWlR87D{uHp7eDtM21M z820!9siVh?{^~*Yi7(JMMm58j43A3M-KF&zL6sj0NS&T>Zvz5{Vq%Ba1t!CUx=gU! z7;+9ZcoG?kQpkPT7dKbkf&lVdMRCvy&j)o{i$0vS=#HF=#vCfT>()!P0g|8wy^Y+f zSm1Xg49`*lsS1Rk2~n{?TH-4Sas=B-Ph%s7Dv@`nvW*~ui==wsn$16$i^}^^`q zH;a1ngNy13j2-o|7w6qW1sMF`qJpD;%pl`Q1HV_kE-4l@V(i;R?PG$+V);I_sLBmV zEo!Qn|Gua^7#uPW3l~+9TS}y9)?j84_Zo2eY9S)1W=dWIP26euq;Hiz-rS9ydgaY- za3j=~aEGuPMclgqo`AxrEDZ9hsXAI_>5VP5W{}&e<-<0=QX2%Nw9l$(Q7Z7(ViL&E$*%k0TpIBA!GFfnPcQ>EHC zK=|#gPC*Z2i|@D&bVXxDVn>D_2Iy@%m|0rE3H+vr@3)ZvOBw`F*CjIe4#%0oZ7Y09 z`97}>64tjn*#F0gkBmAjhO9iQ04TexDlBCxa>J#jO^>-1Eri$f!LU#dSo2e-gk3FMqotVkMONAfB<7;7rVOtWP^Xa;hDCeQQ`K46NmiAG z{Xg0i4`XXf*Y<_M%1kRA^^HI+8&x!Z#y@F5nPLM*dOZ(rkJ?B)rQ}p6Kh4!uB;l z$g9_HnKi4@lD8zDNnC%mbC+E{1&S@^Xaz5o*Om#%#GXm${bI@rlnVL(*N5bV%EZe& zSL~fK`cwVtGKvXlLZvB9(VaIC$Q$qfdT@A$G%sHCa#N?IFH|K`m?;=oH9&M5tcKGDB8! zmsKt2F2hgJi&69^?8Nt0F~z9xBd!oNQ)MOmUngX=;o!N9m ze36lc7&kT*1Dwz8+dqkN$E>Nprtmh3w)b@Hug0sgdeNCcmr*}mQIdPC-~-0=i{Q6p z9?+C5hOA-SwMX;@+q|8e$rm@NKWvN}svi2zRQ@zw?C2ys$#&Lmf-#lK+)_tU5IN`r za6c*>Z{8Zyl!mF?^jl@Xj_)CVb;M^U-cL0)GgW`R@P6P?_cxMqhYZs}ErcUaW~b_p zaFXRmqaRCHkWr14t|ru34EuGQ-9+NrMqlIfq~(mA`T#FuqA`jjNKYMt!$$61|j^Z&DY^VU=&vC}mEneUo0geUtdfpA^1HS+f0x zGdB8M5S_+5eSLpXwjFeW{yC-HVK-~y~Z%bDzQv7jKwcWf3P!78^*Ab1n^rgEn9q6<3Q#x-x ztANo_OZR+5OWx2OGEpy6DZ380qm)JR=3i|bDkO^HCFo{7V~vXbfJVvTDrL3jy~|G3 zpVL^MCDQj^Mfrax1f=Uy9Dm~4x+e5h97SCI+}5e0`pxD zDp_Ted?%M!w+eDywy@&4VrcHyZ#H?EPydxycH*Q!^E2pS9!@YLd_8pg}s}%iA=4Q_+PP6jI*d#|9a7M&3R1069jkc$|%2)1av$gDr@V#hctY zTC>sL8=K&RG084~SPTj#kZ}CcPO4~0S>WLA z9?B98?YPeMR<}5HJxg9w9DzmBVKiu-tN9aUC>2QxLy6d4kh}_H)MweXJeKOq@6pey z3TY_Irn#kmhfG>qAX2TkX~=_s$@IC%95Yh6gJW#8F?>pQzf=scs^WpBqEn3ev?kpQ z3HT~Lx~=gMtI~0SpYwsXi~`H1Q9TW<yVv}-Rsz2pKp!aiW-L9Vx{cJ|Vv$KdFV&HJ z{f7F?ORM;`@=B)`Pd$neLUIPI?KLL|!Z z3uOQ_w+(;Op>`djSBDA`zf{#CcVRJRx#y@!+Eml1d-k89E#h<>&AZbnLcHNn_lF9V zYF8Ex>ce5iE=f6~dhW(5c&(dH?9DGxiu7yB11dZi_9S8EdQxkfXjuoQIf>s~Ncizj z?HPa^J@a#?GYj&1pBMd}%0FM^MLOnpgFq7)_s^q#L;bokoJ6j_-ZxE=dM%qo$25?h z_MQzf;`{bsX-(%hOSG4QeL@dy%B*9id3EaKbf~bB3lP(2W$PPlk%#zMz@_Ttuo8$d~-04WYD!i>Asfo;#O?=5gHJk ziS)shPqEjr&84+Dhg{Y>3-66%3@`QRmfb+1WoZ3Gjn$=booa6>7*a#VOv>|>WM74P8O<$Bad z5@ly1kUb#!P0iqKH@$nj9U3DcmYvP@?vjhKJyvV za|Uu0tHe-Qt-gqtsG^)QZb;(O^Cs$CA_;(m!$rU#sF1c}E4HAqbLnc;-2;}n_;{x=SG>W{l{xQU2b{KVf;sFNsP8v58{t>Cs=f{R0Xg^QYw7yp z(HYQTo9m9Pr_9qIbizt!{qZ}T!>nM6@_63VUr%7LqU!wmhRDY6Rb_~sNX_p=Dt&8( zP%fLpJB317C-@mGT}!q%b{dH$*HYO_IyuRZ#Mxx66dg{ zkiHKbim5qw%tv{3>bJH7+$E@CWwXa#-5sj_Fhoi}<%!RBdaF|*MMs`8=A`=xUbX5m zisk%S{A6J1-OrpV;Jm59oPMh`jR&+n#yC1_^{XaOb5B@K1APy!|4MsN#!OU7_2kt} zR#BOq(R%+uDR)Mb?bsf>dwaCH{-&DNTPI(=bGK7W(Mh-6+jh6+pUQl;t<3}i6+-aR zDg%AUsKGL@N?gG$<@R!C>AP(9+->Q{nSp&z{+QEy>a2N6opr`xL>bXQ16l_1;^sj8 zF%Tsg8KfT>)fR;nO5LTT7)ISSF~kIdc&j)woxCf7(Ou}WfgvC0;$1Exsp46le%RA{Yi1>!@LPRy-NQDL|k@62cl?tHpFL-M0j|K3-r zUt_&e4C7z>I+3z3sgPN{x3=MCa%J3M7cm3xo7R zx%c|=PwuPK+rIuexfFfr$;Bj3PCp7Z0_tR0>3;`^ChX0Bc~X_tTdDuy6j`PJw^IZ( z`1^DCPY4Yl&LEEB=+4&WV=ml3DR{9dW$rlO?ee?wfPGu*2+HppH#u+mMU764taO**}Ps~>fL1T0q zG+WB*lgiN;|Ij0bpapprdpfHUs)Q)bgf~AHu@x*4zfKd&FAK+1xPa}QI)h#rF%>fy zABZ@5zxb4A7#+WcjBov-=nR^|kdtxjt7CXI-@Ri&&^IewOUld{CI)F^7txquDX-Cc zN#%YSw;b#~t=|72qqU2%>US{aX>*1pI>S)1Giezs&$|XvfH<@JLl~W%m zJDFT|(v;odN4&HYNzG41-i+S5KUmqdnn!w)v2-T$DO4qSk~E0adRpTMb27!xe?q)pl zSJi5oKSd+*9^zJBYy5A@`d9oI~xpjbZ5 z&LBm~+bwv&QRMhhSLHsCdNIG@k%?cT7vuWIV=@J5-84Z_MNu5R3ygo6$Penipr5CU zq3IH4f)DlJ=i?Kjhtm2ipK6>;!=sYnwA} zV7qZj)0Sh5$E>KOhe-@x?P{+;%Y5HjDq}W7Vksx&0Pn)4DG{zL(OerNSlT0{!Ykv{ zhBPz%lb^e;pgvJ%FG2@bIt(?K&Tlb$)62KUwj=yj-n5%&fRRubY1yFOV#>|2Nz2x} z4~p@FzUg_F#!+VAORMEr-;t=0y83Y%xNIOsN&O>Bhf{L3!qf~X`>_f`9F{{hL!Kj* zijYU8knK!3X1=={7jRSC$?Bxi+Q9V#qcP}^$_n8tdNcE&i%$BCTP4qSqtAHeK^L5= zJ_&8wypv8~`H0HNbe6u%(g%P&wSKE0s6h|(P?xz*eB z3%tv3EH~noCjNXMd(&c;``fqVt{>3AR(!pZEpzN%GQ}S?PRfNAFHE4MteUo>5*Zm; z?=C<8yN+K@R!*@YG}eCosMN0<48Iect_V$};*jL$2S^!Tm|0Bla@Fn9T03@;(+J`U z?4w=xG=@>@w2&TP!mE?}nZh>LYo_Ccd6_YbhR!UiybS*|VN2;?VoYzGdhyg+0dLR0 zw#rL=HQ1?^(8>F_*?xa%1@(>R;SM^Wz6mEqwcg(t{YaJV7o6_2TJP2-4r^hmOEP7r zZ1NuYeeb?y?B-Jn(WPF1(NB72oYa4#T#ZpHR8DIH4+Ce3e8j=7Va50f3W|VkX3Zn5 znfbb49Mx&RA3X|ZqBzmTozZ8s)fDC8KUSu{1KoO8^P5!4sD3t2rb<{QaxgYT{#o(C z%EN<^mNL*Yr1NA$YzCfo`jIkO8d55#o!5 zclP4flb$&BL-~gOt&YKTiF!dlp4WE(KI`U9;UYQ@aWnR+wLilM92Mwqr0k)+x9q0k zr<-xpPjvLr{m9o-rciKmd=n!Gy`G{z@N7$EK3b5N8`vu$TO|ERgpSiI&wTkY*0Hcn z1;;2nj(XD^sZvXk>r=p3!3#s3TE)DUAG8zW^5RY(!#GC24j;g_zheNsCEFX3pB+PI zh+okHxc+{0QU0>;>!1QOA$N6Sz0~&t zbtf}u!n^?2oKWi93Negv=g<3YX~!AoP3lwrIoo7;-U<&5k| zrV~{Y1HJVZ@|0hc`qF^;qTK=-Wa%I~%4N!LTn-h8D8!r?MRvYxnL;X${-$orS~t!8 z*0R=VOlv7vP+}o1Vo=frN<8dtQ(MM;bsIB~d+1xb40YBYvetFCsr)gNfQe!W(~_dZ z^&m9IiBs?8Ry6$6s8s0ji+s@1NyhQ38lFeWEp*n}#IIQxf(_(SeJU9~xAD_q z_4v`v2I?%8=I3`$obQk73_Q3ct{y`fJ~R=5TW@bnm|Mn3AMV(N6pYc^K@6nNl>YLs zvPYcfTa`emJRuZWR+b*zTN$}{1HT^a9*W(#LFnk98-BXPO4ChsMU<)LHLEaB2R19C zKQ=7tGWnua0fPN;nx}NEa1pZzy8et#lK`~Tb^775ugPzt|9`z*2Y6M*w%#Xjl0y!L zk{r4xkkCSnp(rhv7EnN>J-Clc7aO27gIquiNZ+Ac382&vluHxkDuO|Js3Kj!kU)SS zl3)xFN+_AP{Hd&(I-l4N zZ(IgbCLmw<9|EWiuijK&78T}q%8cgsHX(?^GTwP_J|t6dpRWP zn@4;4NUg%814#r%m!uO>(R&kHi~=ysIH~|N>)HLb(jEI=_%XkLRKXJjglr-p{``{y za$8Qd+R_M~^gT@}0QqQmyb68xzHj5!_;p4#z&_Z34;X;3SnriDQnBsllg+8pucw~Crech)H?<6mhVA~%7FH(i$guZ1toDWL)>w0QAI<%r-An;4RK?4XP?G~ z5mK(U4s+PGP0MQ(_`g*KGA^oCi_<~X-jJBS(Pg7r3%;pJn%Ia|2A}SwDKT1y=_2(D zR(vK>ze;e{*pLtK20*Y$R}0lf)K8$BAY$WL`2j__g+x}PE{f#9AQs7UpMl7Q z(79r0@$sxBJ_}-rPvh+#La{V*=xHH!EsQ2QTo8yYB+|M9N2(DZSkb`StfAYSxtC+P`381027+{!&P@J^CbX}cXiKXwlIlqvJn*NHR7KCao@;rL zRO``f2v>#mV}7g~+a#SI_1?V835k3=V;`4u;O65Co=VI09unG^#WB+d$2`)ngn@$Fcrflv>u9U7MpOf)1zx^*_R4K>$jvb)~i z#}UU5I_wf`0|w&(o1)3EV5TOM^g7!~6D`^|N8!#1+{dDiqJ%3{;8*W*)IV&%XqrcC zMGuXI>SMcU{aI`1Pm&$h7<^e*Vq`g~GO@5^{Yh)qpHyFh82mi?-@0uVtwn#aXP`gp zDq6Gtb9!p6Z|r|v>zY@8leH%OhcUh(qwTT&%pocIcN=czw!%TpHU?yW(*G|emq2p7 z3BC^`GLZg}mg>f~#?!Ux>Q47CQJr4J`=Yw`pOfnH%AmT$K|yt>rd6jG-a%AHBWTs7 zYe03Ooq_6B7pkw<<_sC?kNtS44$(w)IfLr-;(oAHmlw*fmbn58s;(Xf71yK&4?0Y= zRHx3CUqEYa3DVl?@Itr2U>;s)=M?Vnf>g;J;z8STR)-g@ z(cz^V=?;&tsWIaNOebF;GOZUGUPc!HhxLkTX_wqs!~3@s?FDzj51Ac&4tczl^PXj; z!Pmtoli4Lw=o~t!Y7M`2HQWHpIQvacNQ-0xEsN1ap z3`N{RP*wHVzNEWiw&(W$ z6zM|ZF4eVP%sn&tCFC0eMr*P@ZC#dyxmS0(qp00FzH89yPQE0dJSgo%fWl#b{{_I z1igpWz%L?Srj&nsEJ!aw zUwYbDAklSSg6LXk9?A^ARbVLR_N6@I;a&_o@*;sE5rLqsfJrkE1qJglC!Q_%S+w*p z;sHvUj=&;WD{;V=tsHOf1XhTJ?*N|WpAYJQLQs7IIqZeD( z+#G9G#`+1)`GQ5jC10r6ub)tCr`&52Aj8EjFCetnpAh;FN_YDqRrhsim=Is7E5w(V zTylvze&1IYDqHZR!!Dt?R(G$^Qf-(cl(D8w=F%b|k`cK4Lu*=sRfWSP={DJ1v4}mN z3@E&yNbnlX{5!SM7gb)wra?uJEWgr+dQ|AGcjyKz&&oWeLR%m^DCBcws{?YOk883< zxMdeXW9;E>gYCK!5#k#hVU|?a$C4C5*z|w=4-jOQL>d+|mu6mg0B|j&74`WE3%k?s zJAuGd$W(u;*Ha%7kO1oQ&yC86*!UC4I-%k(mNz^Cn0C#O5bBEs50%b_S==g`xt6{+ z+XXhh2MCT{q}2SIzzZ96g47Jk!9GY3e1c(!wLTALc(9X<{{Ap#-iNgV@-B68c#2!8 zE>{O2QZe-&+1FpLi&E;C_jS4flp)u_(6)yfQ=I`C-swrSmR^Mjas~`80dw{cK4op~ z)MxqvJia@`V1H$?GSf+ljb#e_1>=IS5OIWw{NxExQ^=dxU_92a>^$frD zZKW}dlGylb2IKKQ7Krg(tCPeZrtPLCiibNb!HF#XL(fz3H8#TSE}!agm*ArK`E~6d z)$8)QCNFXwh1bpH7i*fNw}4JpaMcWK4{O!z<`pt6R2IXL_r%pO7R<65h8yy=gLmbc zXEkhjFMt9r-iuJ=A)QD~=4CRkkek3A@dh=@8^E2loD|}?fTMzK*Nz7`c);j5)s3B1 zjvipwc&$pnfrzb8$(|pPM~C=43eX{JG7VXAY$gZzI4tmwLIFs~>zuy;ki&^(ifDMH z3}ixHFOk9<5!*p6R34~FwxDEc;7#e%=rzpEUP8ho`B5j$KuXU(jE~r7O7L3kh;61a zjv~69UL4>7yTH%3HX~7~D7?v@rKCl5(>N5;oukdI^6SofXn>}?)4pwuyb(6a<9isaOl3h?*-5!G8FUO9~1PveF+Tk z+P&UmK!P{+3Z{T=WI6T^YROg!BrzTNn<7QngRPB;%btyiLQOI12~CkX2F!}LQkNhz zqnO^Sv6lnk+*NPj&Y1%j2XMe>i*m#gUFo7wx#`?SqM= z6-V&T-7(fe%#~e$JUa)IP&q%Z_YMl|10-1c0G240Pk+Cz!T{<+&X}r6F3%oB)55-b zDW{p%@F{9G^++~3(C=PW(I*pl+c?-fAEL1a9C;YylV_&LyM>)QiAYmeGl|LNICz4( z7vtg^AI3-(^toWM+=<_-gt5v4yDI5N+AQ2c1u))>j;WhfGw=S-X8en;@j&#-P zWsi|$_}17u^ui9TKO$s#gQe4y#Y)wfwiW^YG&$^?w+h738KBFdnKuI@6L-2d)S%sW zpGya5(lN#jQ!1y~XMn0Z0(Q%`Is+>k{G-KzLX&-9qeg_y>KK8}ukuMVRN+EfwGGNpLv9Hl;ZVEzK z`Y0m_z@=ZpndF~Z4(W$;u#EbSuG)rj_F`9mZ0$PkQr&Cm8~ioaG1peMLyFV4I7502 zQXuc$o+{VS{>C6hMLmv1+> z0vnhc(7XQ)Tm09Z%^5@mq=7I%`yi#ORODoAOh=?rZ!Fc222d#kn=#6ZE9@P_#eml; zV;@=%Ja%dH<&Zr0B zH`jqFk_z-m4Yn%qB))`X z%cKtY5Se5fsU#q%RQ#xpXY7g1@_R#FwcE@6y9bufKqz@(9}U z(>jPz-QRzQQb&Maf8+@8@oOU0uvbYMVrLQU)C(p=g9f0aY!cfFcUD!#-w*}N|8^lx z$ciAPdq!bnoC71K(2VY(*p(C{SOgA?_~kk&g&J^tVn0*Rqg8`C(O6C*M%wr+X#+x> zlY36rz$+vcZ2q;NkK`|StAR%f+;L|w-!ay%Y<~x5?u?_@bM?X0@(`hX%wurBf*GkRuL8Ea~$(unpC^ z-I^mnfntrWnam_*YCb8W?kSw1TXa6BG@n+D+EOA|Vz>`90DEKC?Le*oL6=b^ElMOa zih%?IzIo~c?KcKsndkUmm*`~p{I(Q&E&yM8fK02BU+MU*uVR`o!{_o`(7Pm2p?i_w zxCv(Q-k0YoqjMFYF>w+m@)BWuk$gV2JS_(xT4(bNl{55V5Vg3?{d0A;^DrSTHC4_f z|1s3!oFhSVBj0g*jNCypZe_kB%+Bza%80!>F!!x}bc;J#?t6WkUjO{=_vMah)*Zg# ze8)I+j@ga&Dr0mqkT{wl2(6jLb)kl=enY8IVBA2kno$Q?v~^Oz>b_|qtxvk;Y@r)( z4Ohm8pknW`Y$UZb=8hnMzf!LuE0N|hBY;*~MxAT5nTFYcjMefqlb#-H9}0}qlNyeU z$>*>2e_x}k?oHlAhj+?oB){bVsc7=cb2QesfE9LKa4BUAtB%mC*?v91O%tQkDUd9m z3YbEaWV`-NIfM0XGvgT+OLy>VaaA5=uegJ}pzlc$pfZ(Qt>6vI{ zkaXH)Sra{iF2{ioBhF5WM4@nG<# zRlMYl0S{L~^(efHt6-M9aVks)K$1;Px-}0$=U>~E^!%2Vyhb}jJ@&jiLm%3OXIs}c zeoQs>$h=B5v5V@vE@WGKfkSxxL<9qUbNM*^)DL0Oi!giP%y<*4q}~2u7B9u)=EIaR zlB`a(3Q$ykgIjk_y+s@c7C7j**4!992jRwjK_>~-=_9k6^*5yfRNTv_r2$m*kco8h z*r0g``EQpd4WN=HHq;GZ1OqpV8^EaG^@=nAqbG9%7(v_X!@BIbxh0f+P_O~ix4RPL znp)e@i;UIvE?PpUc#JqS04X#?H-Pa{CFDG=M7|4RywpmBNpRp-I4FQ6*VpTCxHN## z$7~}FV2nJxn;XE5?lfX%yi12?Nchf$F#s(MZ8b|$sf}Q2IF(TY02%Z(b%49)ltjv@ zbeILZ{TtAs{COQc7XU5GkE)9pCRdZPcz!#5eMw3xb0%k6V+k_%BIdPfz%gL+6L1yk zWN|>+(1=n8jwaHt^tg&g)<`|fuEyKH`oVBwxcX;D&;bUcuh%?U#P|BKC5w-Z=fSjs zC* zU1HHv20_(h~_cbeapzd%5$nFsg?+U8S2q1_uv0lYCG+HYC;|4x5 z8;)pKI*2M%Do6+`eHNr!Nvq*zOmBgZ+p#l%6HbuJX5S%Cav^Zi!`?T^!jOd6QIcOgl!gr{sKx&RP z__%`ehwQ3wAE&8d#yK*c8iq6F8rq$?c|hJe%oGFH$-mN)iwSR>9y73cUiVObj^Y2; z2xY2q4Axo$K(n}F3r#UITin6nw*GU)mgD#{OCNB}7=ZCT2Hx^9+djjev*-Y$k*)K*?qJRx37VfK0roTXQwHe{VW<~c7QjaJY5(XyX} z#%a|9KpfHJc>w)#7LFH)=Dd1Lz@l@zltxeX-UT#Q^hKjJJm+C86HN3j&o0K3c02z4 z{S_xiq=|!YypQvL0KH>D9{_lqV_;xlWB`Jbo%x0!`poS$axfYIM;Qjz004NLV_;-p zVBiB{Ed~%^VqjrlWM~D_Z4B)|RtLibAbTRi8X%hmEYAP{Rlx)_000010000A02}}w z00C@abZ!6$00aO40096000000004NLeO60z6jc=dx+jmxh!JCmh!`WjV?>M^F(M*j z#E7UE5fPEZ2O&m$p`s!xVoDLwWTxkt9wwcsB%Th z$fPQhN>%3G^PPLoIo~~xh65mmR&>A_?mu)8QPfa^ix**tGkjowKV>NGh}Gyf3MV#; z7Ifmg)8=$K8=bt{i+y^KK&0!2HxKDGPpoX3MmV2M5$Gu5v zP{;oX-jJg7JKD$-Pf+G?J%+f}H&D;V1XuSXLQOjtLmRrV8JBbA5l!>HNm};h$Fc>| zB7Ms9?<(FI=Q>Ah8}INa8x>a@=tn@^Jn;!hT4~9{Y04XEsS}%cVi%5*&wLnrPBA`9 z$qfA;li%j!ah{nW9+&lJXtNxBJJ%(j#&9F=pQf|{&6IDzcI?GL9Ki^=cIkaeaTv2H zv5><$TJpK~6{W@S3MFy+n2_H-l31?y$nl&-nZpV0J^oJdT%11D@h?ry)ADB@<4g(t zoU1r>a`=LN*~+D6E1tk*yuGMc#$~GNa-KhM#ZlEhvm%S@<~jZsRc z@v4YiDg0?u%JtQ)kFr#<+DnrCqy|2mCMyVK(gI7gkVbJjWM{O#7bt?qhvhdbmRbF=Pm z?wtFl`(LCk(i~aCbB#QiR((t=1G1`tamD!I5~#=As*WaErPa|4-dqALVE+=Rk9Sxf z^QcwE1i8!V-Xm49B44{2R=}aESP}=SVm|Yy8s_17VtGjt-2bFEwEQYEOFw1~%%8K& zDTZDSRY3~)pbFyQttvgOY~?H+WR8sxFW?t;@ZWI`f8YZCVGnCizS+BF!b#UJ zkKz^GeJgF0d$C6jJH(qYoD+{H)z|Gr>1*Th1D1BYzakqN^O!&ppL7nTlN^8mm6%601Z-YQD@<7<>t#d}h zITLVa{nftzx*7d_L=rlkRfI`=&zMFv)y(I;+Lh<&gWM%9@%6){onaB)w3-wAb111K>2S6?d16UdWGrS8W z004NLV_;-pV0`fZ0fQbB0|Uc91|SI(K>?EhZgvEP004NLoslgL!axv(mrZJFFf}!p znkq~~Qo{|_;?S5T7%qS#=?%~rE`jD+oeAGdCY3b#vTtYS=WREqlyXW#{;jPsqMaP{ zoxBcqp!$LOgHf$wy|FQB@;l?Iw#{1d=BJ06ZFiT z@ino^i{?su|4W!vIIC9r81CTpFXP6)^FCX0L}9LY;%P>|gJ;*Da4xNPQ{4Mpjc>TW z-oQDt)_7Lj{q6DSwCV3}0b{rw1ONa4Pyhg}004NLZH_-~!Y~|#zd#5=McvxjQq?h$ z?Fn&!9H0X`gB2m61eq{!kxD&5H;%ypdV(IM&)}g0wygL0>HYoyOFk(`r=a3F^MpC? znWwyQ%3Lb*oq5I+*UXjfuaZT%kh8p1C7!6`OXevpb>>o;ugo)Exn!>7{}AXIIdfou z(2wVX^h0*1MfPlo`iAbb-wIlVQaou?B4F?|Sm|2tf3v&eS!ofcTZ^sKtIg1l?GU?Q zd}Gbp+{X(1s&=mS)+zLGv5oA=k=<<(qjqUHn-Gt?p*OCvj|=|*pCUI_004NLV_;^$ J3jhFp00ga`WU&AM literal 0 HcmV?d00001 diff --git a/app/assets/stylesheets/icons.css.erb b/app/assets/stylesheets/icons.css.erb new file mode 100644 index 00000000..d71b7cf9 --- /dev/null +++ b/app/assets/stylesheets/icons.css.erb @@ -0,0 +1,519 @@ +/* WARNING: Your are using ligatures for your icon font. + Ligatures are not supported in IE 9 (and older). + Use the Private Use Area encoding for best browser support. +==================================================================== */ +@font-face { + font-family: 'entypo'; + src:url(<%= asset_path 'entypo.eot' %>); + src:url(<%= asset_path 'entypo.eot?#iefix' %>) format('embedded-opentype'), + url(<%= asset_path 'entypo.svg#entypo' %>) format('svg'), + url(<%= asset_path 'entypo.woff' %>) format('woff'), + url(<%= asset_path 'entypo.ttf' %>) format('truetype'); + font-weight: normal; + font-style: normal; +} + +/* Use the following CSS code if you want to use data attributes for inserting your icons */ +[data-icon]:before { + font-family: 'entypo'; + content: attr(data-icon); + speak: none; + /* Enable Ligatures */ + -webkit-font-feature-settings:"liga","dlig"; + -moz-font-feature-settings:"liga=1, dlig=1"; + -moz-font-feature-settings:"liga","dlig"; + -ms-font-feature-settings:"liga","dlig"; + -o-font-feature-settings:"liga","dlig"; + font-feature-settings:"liga","dlig"; + text-rendering:optimizeLegibility; + font-weight: normal; + -webkit-font-smoothing: antialiased; +} + +/* Use the following CSS code if you want to have a class per icon */ +[class^="icons-"]:before, [class*=" icons-"]:before { + font-size: 1.5em; + font-family: 'entypo'; + font-style: normal; + speak: none; + /* Enable Ligatures */ + -webkit-font-feature-settings:"liga","dlig"; + -moz-font-feature-settings:"liga=1, dlig=1"; + -moz-font-feature-settings:"liga","dlig"; + -ms-font-feature-settings:"liga","dlig"; + -o-font-feature-settings:"liga","dlig"; + font-feature-settings:"liga","dlig"; + text-rendering:optimizeLegibility; + font-weight: normal; + -webkit-font-smoothing: antialiased; + display: inline-block; + text-decoration: inherit; +} +a [class^="icons-"], a [class*=" icons-"] { + display: inline-block; + text-decoration: inherit; +} +/* makes the font 33% larger relative to the icon container */ +.icons-large:before { + vertical-align: top; + font-size: 1.3333333333333333em; +} +.btn [class^="icons-"], .btn [class*=" icons-"] { + /* keeps button heights with and without icons the same */ + + line-height: .9em; +} +li [class^="icons-"], li [class*=" icons-"] { + display: inline-block; + width: 1.25em; + text-align: center; + margin-right: 10px; +} +li .icons-large[class^="icons-"], li .icons-large[class*=" icons-"] { + /* 1.5 increased font size for icons-large * 1.25 width */ + + width: 1.875em; +} +li[class^="icons-"], li[class*=" icons-"] { + margin-left: 0; + list-style-type: none; +} +li[class^="icons-"]:before, li[class*=" icons-"]:before { + text-indent: -2em; + text-align: center; +} +li[class^="icons-"].icons-large:before, li[class*=" icons-"].icons-large:before { + text-indent: -1.3333333333333333em; +} +.icons-phone:before { + content: "\70\68\6f\6e\65"; +} +.icons-directions:before { + content: "\64\69\72\65\63\74\69\6f\6e\73"; +} +.icons-mail:before { + content: "\6d\61\69\6c"; +} +.icons-pencil:before { + content: "\70\65\6e\63\69\6c"; +} +.icons-paperclip:before { + content: "\70\61\70\65\72\63\6c\69\70"; +} +.icons-drawer:before { + content: "\64\72\61\77\65\72"; +} +.icons-member:before { + content: "\6d\65\6d\62\65\72"; +} +.icons-group:before { + content: "\67\72\6f\75\70"; +} +.icons-addmember:before { + content: "\61\64\64\6d\65\6d\62\65\72"; +} +.icons-location:before { + content: "\6c\6f\63\61\74\69\6f\6e"; +} +.icons-share:before { + content: "\73\68\61\72\65"; +} +.icons-heart-full:before { + content: "\68\65\61\72\74\2d\66\75\6c\6c"; +} +.icons-heart-bare:before { + content: "\68\65\61\72\74\2d\62\61\72\65"; +} +.icons-star-full:before { + content: "\73\74\61\72\2d\66\75\6c\6c"; +} +.icons-star-bare:before { + content: "\73\74\61\72\2d\62\61\72\65"; +} +.icons-discuss:before { + content: "\64\69\73\63\75\73\73"; +} +.icons-comment:before { + content: "\63\6f\6d\6d\65\6e\74"; +} +.icons-quote:before { + content: "\71\75\6f\74\65"; +} +.icons-house:before { + content: "\68\6f\75\73\65"; +} +.icons-search:before { + content: "\73\65\61\72\63\68"; +} +.icons-printer:before { + content: "\70\72\69\6e\74\65\72"; +} +.icons-bell:before { + content: "\62\65\6c\6c"; +} +.icons-link:before { + content: "\6c\69\6e\6b"; +} +.icons-flag:before { + content: "\66\61\6c\67"; +} +.icons-cog:before { + content: "\63\6f\67"; +} +.icons-tools:before { + content: "\74\6f\6f\6c\73"; +} +.icons-tag:before { + content: "\74\61\67"; +} +.icons-camera:before { + content: "\63\61\6d\65\72\61"; +} +.icons-megaphone:before { + content: "\6d\65\67\61\70\68\6f\6e\65"; +} +.icons-new:before { + content: "\6e\65\77"; +} +.icons-graduation:before { + content: "\67\72\61\64\75\61\74\69\6f\6e"; +} +.icons-books:before { + content: "\62\6f\6f\6b\73"; +} +.icons-page:before { + content: "\70\61\67\65"; +} +.icons-lifebuoy:before { + content: "\6c\69\66\65\62\75\6f\79"; +} +.icons-eye:before { + content: "\65\79\65"; +} +.icons-clock:before { + content: "\63\6c\6f\63\6b"; +} +.icons-calendar:before { + content: "\63\61\6c\65\6e\64\61\72"; +} +.icons-briefcase:before { + content: "\62\72\69\65\66\63\61\73\65"; +} +.icons-gauge:before { + content: "\67\61\75\67\65"; +} +.icons-language:before { + content: "\6c\61\6e\67\75\61\67\65"; +} +.icons-keys:before { + content: "\6b\65\79\73"; +} +.icons-earth:before { + content: "\65\61\72\74\68"; +} +.icons-keyboard:before { + content: "\6b\65\79\62\6f\61\72\64"; +} +.icons-browser:before { + content: "\62\72\6f\77\73\65\72"; +} +.icons-publish:before { + content: "\70\75\62\6c\69\73\68"; +} +.icons-code:before { + content: "\63\6f\64\65"; +} +.icons-light-bulb:before { + content: "\6c\69\67\68\74\2d\62\75\6c\62"; +} +.icons-database:before { + content: "\64\61\74\61\62\61\73\65"; +} +.icons-box:before { + content: "\62\6f\78"; +} +.icons-rss:before { + content: "\72\73\73"; +} +.icons-clipboard:before { + content: "\63\6c\69\70\62\6f\61\72\64"; +} +.icons-cart:before { + content: "\63\61\72\74"; +} +.icons-template:before { + content: "\75\6e\74\69\74\6c\65\64"; +} +.icons-statistics:before { + content: "\73\74\61\74\69\73\74\69\63\73"; +} +.icons-pie:before { + content: "\70\69\65"; +} +.icons-bars:before { + content: "\62\61\72\73"; +} +.icons-graph:before { + content: "\67\72\61\70\68"; +} +.icons-lock:before { + content: "\6c\6f\63\6b"; +} +.icons-unlock:before { + content: "\75\6e\6c\6f\63\6b"; +} +.icons-outlog:before { + content: "\6f\75\74\6c\6f\67"; +} +.icons-inlog:before { + content: "\69\6e\6c\6f\67"; +} +.icons-checkmark:before { + content: "\63\68\65\63\6b\6d\61\72\6b"; +} +.icons-aminus:before { + content: "\61\6d\69\6e\75\73"; +} +.icons-aplus:before { + content: "\61\70\6c\75\73"; +} +.icons-aclose:before { + content: "\61\63\6c\6f\73\65"; +} +.icons-bminus:before { + content: "\62\6d\69\6e\75\73"; +} +.icons-bplus:before { + content: "\62\70\6c\75\73"; +} +.icons-bclose:before { + content: "\62\63\6c\6f\73\65"; +} +.icons-cminus:before { + content: "\63\6d\69\6e\75\73"; +} +.icons-cplus:before { + content: "\63\70\6c\75\73"; +} +.icons-cross:before { + content: "\63\72\6f\73\73"; +} +.icons-blocked:before { + content: "\62\6c\6f\63\6b\65\64"; +} +.icons-information:before { + content: "\69\6e\66\6f\72\6d\61\74\69\6f\6e"; +} +.icons-binfo:before { + content: "\62\69\6e\66\6f"; +} +.icons-question:before { + content: "\71\75\65\73\74\69\6f\6e"; +} +.icons-help:before { + content: "\68\65\6c\70"; +} +.icons-warning:before { + content: "\77\61\72\6e\69\6e\67"; +} +.icons-shuffle:before { + content: "\73\68\75\66\66\6c\65"; +} +.icons-return:before { + content: "\72\65\74\75\72\6e"; +} +.icons-enter:before { + content: "\65\6e\74\65\72"; +} +.icons-exchange:before { + content: "\65\78\63\68\61\6e\67\65"; +} +.icons-loop:before { + content: "\6c\6f\6f\70"; +} +.icons-th-list:before { + content: "\74\68\2d\6c\69\73\74"; +} +.icons-th-large:before { + content: "\74\68\2d\6c\61\72\67\65"; +} +.icons-align-justify:before { + content: "\61\6c\69\67\6e\2d\6a\75\73\74\69\66\79"; +} +.icons-wtext:before { + content: "\77\74\65\78\74"; +} +.icons-btext:before { + content: "\62\74\65\78\74"; +} +.icons-pictures:before { + content: "\70\69\63\74\75\72\65"; +} +.icons-video:before { + content: "\76\69\64\65\6f"; +} +.icons-music:before { + content: "\6d\6f\75\73\65"; +} +.icons-folder:before { + content: "\66\6f\6c\64\65\72"; +} +.icons-archive:before { + content: "\61\72\63\68\69\76\65"; +} +.icons-trash:before { + content: "\74\72\61\73\68"; +} +.icons-upload:before { + content: "\75\70\6c\6f\61\64"; +} +.icons-download:before { + content: "\64\6f\77\6e\6c\6f\61\64"; +} +.icons-disk:before { + content: "\64\69\73\6b"; +} +.icons-bookmark:before { + content: "\62\6f\6f\6b\6d\61\72\6b"; +} +.icons-booma:before { + content: "\62\6f\6f\6d\61"; +} +.icons-resize-enlarge:before { + content: "\72\65\73\69\7a\65\2d\65\6e\6c\61\72\67\65"; +} +.icons-resize-shrink:before { + content: "\72\65\73\69\7a\65\2d\73\68\72\69\6e\6b"; +} +.icons-flow-tree:before { + content: "\66\6c\6f\77\2d\74\72\65\65"; +} +.icons-arrow-left-a:before { + content: "\61\72\72\6f\77\2d\6c\65\66\74\2d\61"; +} +.icons-arrow-bottom-a:before { + content: "\61\72\72\6f\77\2d\62\6f\74\74\6f\6d\2d\61"; +} +.icons-arrow-top-a:before { + content: "\61\72\72\6f\77\2d\74\6f\70\2d\61"; +} +.icons-arrow-right-a:before { + content: "\61\72\72\6f\77\2d\72\69\67\68\74\2d\61"; +} +.icons-arrow-left-b:before { + content: "\61\72\72\6f\77\2d\6c\65\66\74\2d\62"; +} +.icons-arrow-bottom-b:before { + content: "\61\72\72\6f\77\2d\62\6f\74\74\6f\6d\2d\62"; +} +.icons-arrow-top-b:before { + content: "\61\72\72\6f\77\2d\74\6f\70\2d\62"; +} +.icons-arrow-right-b:before { + content: "\61\72\72\6f\77\2d\72\69\67\68\74\2d\62"; +} +.icons-arrow-left-c:before { + content: "\61\72\72\6f\77\2d\6c\65\66\74\2d\63"; +} +.icons-arrow-bottom-c:before { + content: "\61\72\72\6f\77\2d\62\6f\74\74\6f\6d\2d\63"; +} +.icons-arrow-top-c:before { + content: "\61\72\72\6f\77\2d\74\6f\70\2d\63"; +} +.icons-arrow-right-c:before { + content: "\61\72\72\6f\77\2d\72\69\67\68\74\2d\63"; +} +.icons-arrow-left-d:before { + content: "\61\72\72\6f\77\2d\6c\65\66\74\2d\64"; +} +.icons-arrow-bottom-d:before { + content: "\61\72\72\6f\77\2d\62\6f\74\74\6f\6d\2d\64"; +} +.icons-arrow-top-d:before { + content: "\61\72\72\6f\77\2d\74\6f\70\2d\64"; +} +.icons-arrow-right-d:before { + content: "\61\72\72\6f\77\2d\72\69\67\68\74\2d\64"; +} +.icons-arrow-left-e:before { + content: "\61\72\72\6f\77\2d\6c\65\66\74\2d\65"; +} +.icons-arrow-bottom-e:before { + content: "\61\72\72\6f\77\2d\62\6f\74\74\6f\6d\2d\65"; +} +.icons-arrow-top-e:before { + content: "\61\72\72\6f\77\2d\74\6f\70\2d\65"; +} +.icons-arrow-right-e:before { + content: "\61\72\72\6f\77\2d\72\69\67\68\74\2d\65"; +} +.icons-arrow-left-f:before { + content: "\61\72\72\6f\77\2d\6c\65\66\74\2d\66"; +} +.icons-arrow-bottom-f:before { + content: "\61\72\72\6f\77\2d\62\6f\74\74\6f\6d\2d\66"; +} +.icons-arrow-top-f:before { + content: "\61\72\72\6f\77\2d\74\6f\70\2d\66"; +} +.icons-left-f:before { + content: "\61\72\72\6f\77\2d\72\69\67\68\74\2d\66"; +} +.icons-menu:before { + content: "\6d\65\6e\75"; +} +.icons-ellipsis:before { + content: "\65\6c\6c\69\70\73\69\73"; +} +.icons-dots:before { + content: "\64\6f\74\73"; +} +.icons-dot:before { + content: "\64\6f\74"; +} +.icons-like:before { + content: "\6c\69\6b\65"; +} +.icons-suck:before { + content: "\73\75\63\6b"; +} +.icons-export:before { + content: "\65\78\70\6f\72\74"; +} +.icons-vcard:before { + content: "\76\63\61\72\64"; +} +.icons-flow-cascade:before { + content: "\21"; +} +.icons-landscape:before { + content: "\22"; +} +.icons-brush:before { + content: "\62\72\75\73\68"; +} +.icons-palette:before { + content: "\70\61\6c\65\74\74\65"; +} +.icons-desktop:before { + content: "\64\65\73\6b\74\6f\70"; +} +.icons-plane:before { + content: "\70\6c\61\6e\65"; +} +.icons-booklet:before { + content: "\62\6f\6f\6b\6c\65\74"; +} +.icons-update:before { + content: "\75\70\64\61\74\65"; +} +.icons-reload:before { + content: "\72\65\6c\6f\61\64"; +} +.icons-unload:before { + content: "\75\6e\6c\6f\61\64"; +} +.icons-trophy:before { + content: "\74\72\6f\70\68\79"; +} diff --git a/app/assets/stylesheets/new_admin.css.erb b/app/assets/stylesheets/new_admin.css.erb index 8d1e8a8b..a634fb42 100644 --- a/app/assets/stylesheets/new_admin.css.erb +++ b/app/assets/stylesheets/new_admin.css.erb @@ -14,4 +14,5 @@ *= require widgets *= require scroll_style *= require isotope + *= require icons */ diff --git a/app/assets/stylesheets/orbit-bar.css.erb b/app/assets/stylesheets/orbit-bar.css.erb index adde331e..38441260 100644 --- a/app/assets/stylesheets/orbit-bar.css.erb +++ b/app/assets/stylesheets/orbit-bar.css.erb @@ -69,9 +69,6 @@ #orbit-bar .orbit-logo .dropdown-menu { left: -15px; } -#orbit-bar .orbit-logo .dropdown-menu>li>a:hover>i { - background-image: url(<%= asset_path 'icons_pack_white.png' %>); -} #orbit-bar .nav > li { height: 28px; } @@ -79,13 +76,18 @@ background-color: rgba(0,157,220,1); } #orbit-bar .nav > li > a { - background-image: url(<%= asset_path 'orbit-bar.png' %>); + /*background-image: url(<%= asset_path 'orbit-bar.png' %>);*/ background-repeat:no-repeat; display: inline-block; - width: 16px; + width: 17px; height: 16px; - text-indent: -9999px; - padding:6px; + padding: 6px 6px 6px 5px; +} +#orbit-bar .nav > li > a[data-icon]:before { + font-size: 1.5em; + color: #fff; + line-height: 16px; + text-shadow: none; } #orbit-bar .nav > li.search { overflow: hidden; diff --git a/app/assets/stylesheets/style.css.erb b/app/assets/stylesheets/style.css.erb index 6a496370..00ee8e0a 100644 --- a/app/assets/stylesheets/style.css.erb +++ b/app/assets/stylesheets/style.css.erb @@ -50,11 +50,12 @@ color: #FFF; } #main-sidebar #position [class^="icons"] { - background-image: url(<%= asset_path 'icons_pack_white.png' %>); + color: #FFF; } #main-sidebar #position a { display: block; padding-left: 11px; + height: 36px; } #main-sidebar #position #collapse-menu { height: 16px; @@ -91,9 +92,8 @@ margin: 1px 0; } #main-sidebar .nav > li > a [class^="icons-"] { - opacity: .5; - -moz-opacity: .5; - filter:alpha(opacity=5); + float: left; + color: #333; } #main-sidebar .nav > li.active > a [class^="icons-"] { opacity: 1; @@ -103,6 +103,7 @@ #main-sidebar .overview > .nav > li > a { padding: 3px 13px; line-height: 30px; + min-height: 30px; background-color: #e2e2e2; margin-left: -17px; border-top: 1px solid #DBDBDB; @@ -111,6 +112,13 @@ -webkit-box-shadow: inset -5px 0px 15px rgba(0, 0, 0, .18), 0px -1px 0px rgba(0, 0, 0, .1); -moz-box-shadow: inset -5px 0px 15px rgba(0, 0, 0, .18), 0px -1px 0px rgba(0, 0, 0, .1); } +#main-sidebar .overview > .nav > li > a:after { + clear: both; + display: block; + height: 0; + content: ""; + visibility: hidden; +} #main-sidebar .overview > .nav > li.active > a { background-color: white; box-shadow: none; @@ -126,10 +134,7 @@ -moz-box-shadow: inset -5px 0px 15px rgba(0, 0, 0, .15); } #main-sidebar .overview > .nav > li:hover > a [class^="icons-"] { - background-image: url(<%= asset_path 'icons_pack_white.png' %>); - opacity: 1; - -moz-opacity: 1; - filter:alpha(opacity=10); + color: #fff; } #main-sidebar .nav > li > .nav { margin-left: -15px; @@ -265,7 +270,7 @@ #post-body .editor { } #post-body-content { - padding: 8px 0 8px 6px; + padding: 8px 6px; } #post-body-content .middle { width: 100%; @@ -286,7 +291,7 @@ .filter .accordion-heading { border-bottom: none; border-top: none; - border-left: 1px solid #E9E9E9; + border-left: 1px solid rgba(0,0,0,0.07); border-right: none; -moz-border-radius: 0; -webkit-border-radius: 0; @@ -297,7 +302,7 @@ top: 0; } .filter li:last-child .accordion-heading { - border-right: 1px solid #E9E9E9; + border-right: 1px solid rgba(0,0,0,0.07); } .accordion-group .accordion-toggle .caret { border-top-color: #0088CC; @@ -330,15 +335,20 @@ .filters .accordion-inner { border-top: none; padding: 9px 15px 4px; + position: relative; } .filters .filter-clear { - padding: 5px 5px 0; + padding: 5px 10px 0; border-top: 1px solid rgba(0,0,0,0.1); text-align: right; -webkit-box-shadow: inset 0 1px 0px rgba(255, 255, 255, 0.5); -moz-box-shadow: inset 0 1px 0px rgba(255, 255, 255, 0.5); box-shadow: inset 0 1px 0px rgba(255, 255, 255, 0.5); } +.search-results { + width: 840px; + margin: -250px 0 0 -420px; +} #tags { } #tags .tag { @@ -455,20 +465,27 @@ .img-peview { margin-left: 12px; } -/*.popover img { - max-height: 120px; +.popover .arrow { + border-bottom-color: #333; + border-width: 0 10px 10px; + display: none; +} +.popover img { + max-height: 100%; max-width: 100%; } .popover-inner { width: auto; } .popover-title { + display: none; padding: 5px; } .popover-content { + text-align: center; border-radius: 3px; padding: 5px; -}*/ +} .view-mode { margin: 6px 10px 0 0; @@ -476,9 +493,6 @@ .view-mode .btn { margin-bottom: 0; } -.view-mode .btn { - margin-bottom: 0; -} .view-mode i { font-size: 1.2em; line-height: 17px !important; @@ -532,8 +546,13 @@ .folded #main-sidebar .nav > li > a [class^="icons-"] { margin-left: 1px; } +/*.folded #main-sidebar .viewport { + width: 39px; + background-color: rgba(100,100,100,.3) +}*/ .folded #main-sidebar:hover .viewport { width: 198px; + /*background-color: rgba(100,100,100,.3)*/ } .folded #main-sidebar .overview > .nav-list > li { position: relative; @@ -607,6 +626,13 @@ .text-green { color: #39b54a !important; } +#banner_tab li a { + padding-right: 32px; +} +#banner_tab li [class^="icons-"] { + float: right; + margin-top: 8px; +} .adbanner-setup { margin-right: 10px; margin-bottom: 30px !important; @@ -675,796 +701,4 @@ #category_id, #module_app_id { width: auto; -} - -[class^="icons-"] { - display: inline-block; - width: 16px; - height: 16px; - vertical-align: text-top; - background-image: url(<%= asset_path 'icons_pack.png' %>); - background-position: 16px 16px; - background-repeat: no-repeat; - *margin-right: .3em; - margin-right:10px; -} -[class^="icons-"]:last-child { - *margin-left: 0; -} -.icons-white { - background-image: url(<%= asset_path 'icons_pack_white.png' %>); -} -/*1*/ -.icons-pencil { - background-position: 0 0; -} -.icons-brush { - background-position: -32px 0; -} -.icons-pen { - background-position: -64px 0; -} -.icons-brush-large { - background-position: -128px 0; -} -.icons-pen-small { - background-position: -96px 0; -} -.icons-bucket { - background-position: -160px 0; -} -.icons-eye { - background-position: -192px 0; -} -.icons-ban { - background-position: -224px 0; -} -.icons-trash { - background-position: -256px 0; -} -.icons-zoom { - background-position: -288px 0; -} -.icons-zoom-out { - background-position: -320px 0; -} -.icons-zoom-in { - background-position: -352px 0; -} -.icons-magic { - background-position: -384px 0; -} -.icons-aim { - background-position: -416px 0; -} -/*2*/ -.icons-flag { - background-position: 0 -32px; -} -.icons-paperclip { - background-position: -32px -32px; -} -.icons-share { - background-position: -64px -32px; -} -.icons-link { - background-position: -96px -32px; -} -.icons-tag { - background-position: -128px -32px; -} -.icons-lock { - background-position: -160px -32px; -} -.icons-unlock { - background-position: -192px -32px; -} -.icons-thumbtack { - background-position: -224px -32px; -} -.icons-pin { - background-position: -257px -32px; -} -.icons-shield { - background-position: -288px -32px; -} -.icons-key { - background-position: -320px -32px; -} -.icons-fire { - background-position: -352px -32px; -} -.icons-bulls-eye { - background-position: -384px -32px; -} -.icons-flash { - background-position: -416px -32px; -} -.icons-time { - background-position: -448px -32px; -} -.icons-halo { - background-position: -480px -32px; -} -.icons-hourglass { - background-position: -513px -32px; -} -.icons-alarm-clock { - background-position: -545px -32px; -} -.icons-paper { - background-position: -577px -32px; -} -.icons-banner { - background-position: -608px -32px; -} -/*3*/ -.icons-phone { - background-position: 0px -64px; -} -.icons-mobile { - background-position: -32px -64px; -} -.icons-mail { - background-position: -64px -64px; -} -.icons-mail-open { - background-position: -96px -64px; -} -.icons-mail-read { - background-position: -128px -64px; -} -.icons-content { - background-position: -160px -64px; -} -.icons-content-out { - background-position: -192px -64px; -} -.icons-content-in { - background-position: -224px -64px; -} -.icons-projector { - background-position: -256px -64px; -} -.icons-tape { - background-position: -288px -64px; -} -.icons-chat-a { - background-position: -320px -64px; -} -.icons-chat-b { - background-position: -352px -64px; -} -.icons-chat-c { - background-position: -384px -64px; -} -.icons-comment { - background-position: -416px -64px; -} -.icons-rss { - background-position: -448px -64px; -} -.icons-ship { - background-position: -480px -64px; -} -.icons-send { - background-position: -512px -64px; -} -.icons-bell { - background-position: -544px -64px; -} -.icons-announcement { - background-position: -576px -64px; -} -/*4*/ -.icons-contact { - background-position: 0 -96px; -} -.icons-roll { - background-position: -32px -96px; -} -.icons-member { - background-position: -288px -96px; -} -.icons-member-user { - background-position: -64px -96px; -} -.icons-member-admin { - background-position: -96px -96px; -} -.icons-member-manager{ - background-position: -128px -96px; -} -.icons-member-plus{ - background-position: -160px -96px; -} -.icons-member-minus{ - background-position: -192px -96px; -} -.icons-member-blockade{ - background-position: -224px -96px; -} -.icons-carte { - background-position: -256px -96px; -} -.icons-building { - background-position: -320px -96px; -} -.icons-calendar { - background-position: -352px -96px; -} -.icons-calendars { - background-position: -384px -96px; -} -.icons-out { - background-position: -416px -96px; -} -.icons-desktop { - background-position: -448px -96px; -} -/*5*/ -.icons-page-blank { - background-position: 0px -128px; -} -.icons-page { - background-position: -32px -128px; -} -.icons-page-copy { - background-position: -64px -128px; -} -.icons-folder { - background-position: -96px -128px; -} -.icons-folder-open { - background-position: -128px -128px; -} -.icons-folder-lock { - background-position: -160px -128px; -} -.icons-folder-plus { - background-position: -192px -128px; -} -.icons-folder-minus { - background-position: -224px -128px; -} -.icons-page-plus { - background-position: -256px -128px; -} -.icons-page-minus { - background-position: -288px -128px; -} -.icons-page-edit { - background-position: -320px -128px; -} -.icons-page-download { - background-position: -352px -128px; -} -/*6*/ -.icons-house-w { - background-position: 0px -160px; -} -.icons-house-b { - background-position: -32px -160px; -} -.icons-signs { - background-position: -64px -160px; -} -.icons-globe { - background-position: -96px -160px; -} -.icons-map { - background-position: -128px -160px; -} -.icons-markers { - background-position: -160px -160px; -} -.icons-barrier { - background-position: -192px -160px; -} -.icons-assist { - background-position: -224px -160px; -} -.icons-cones { - background-position: -256px -160px; -} -.icons-group { - background-position: -288px -160px; -} -.icons-cuble { - background-position: -320px -160px; -} -.icons-structure { - background-position: -352px -160px; -} -.icons-layer { - background-position: -384px -160px; -} -/*7*/ -.icons-shopcar-a { - background-position: 0px -192px; -} -.icons-shopcar-b { - background-position: -34px -192px; -} -.icons-purchase { - background-position: -64px -192px; -} -.icons-shopcart { - background-position: -96px -192px; -} -.icons-van { - background-position: -128px -192px; -} -.icons-form { - background-position: -160px -192px; -} -.icons-gift { - background-position: -192px -192px; -} -.icons-credit-card { - background-position: -224px -192px; -} -.icons-cash { - background-position: -256px -192px; -} -.icons-assets { - background-position: -288px -192px; -} -.icons-computer { - background-position: -320px -192px; -} -.icons-library { - background-position: -352px -192px; -} -/*8*/ -.icons-dashboard { - background-position: 0 -224px; -} -.icons-cog { - background-position: -32px -224px; -} -.icons-cogs { - background-position: -64px -224px; -} -.icons-tool { - background-position: -96px -224px; -} -.icons-screwdriver { - background-position: -128px -224px; -} -.icons-wrench { - background-position: -160px -224px; -} -.icons-toolbox { - background-position: -192px -224px; -} -.icons-switch { - background-position: -224px -224px; -} -.icons-valve { - background-position: -256px -224px; -} -/*9*/ -.icons-book-cover { - background-position: 0px -256px; -} -.icons-book-make { - background-position: -32px -256px; -} -.icons-binder { - background-position: -64px -256px; -} -.icons-album { - background-position: -96px -256px; -} -.icons-camera { - background-position: -128px -256px; -} -.icons-video-camera { - background-position: -160px -256px; -} -.icons-pillar { - background-position: -192px -256px; -} -.icons-chart { - background-position: -224px -256px; -} -.icons-picture { - background-position: -256px -256px; -} -.icons-pictures { - background-position: -288px -256px; -} -.icons-brief { - background-position: -320px -256px; -} -.icons-film { - background-position: -352px -256px; -} -.icons-asset { - background-position: -384px -256px; -} -.icons-asset-download { - background-position: -416px -256px; -} -.icons-asset-upload { - background-position: -448px -256px; -} -.icons-music { - background-position: -480px -256px; -} -.icons-book-open-w { - background-position: -512px -256px; -} -.icons-book-open-b { - background-position: -544px -256px; -} -.icons-clapper-board { - background-position: -576px -256px; -} -/*10*/ -.icons-date { - background-position: 0px -288px; -} -.icons-screen { - background-position: -32px -288px; -} -.icons-iphone { - background-position: -64px -288px; -} -.icons-ipad { - background-position: -96px -288px; -} -.icons-ipod { - background-position: -128px -288px; -} -.icons-battery-low { - background-position: -160px -288px; -} -.icons-battery-mid { - background-position: -192px -288px; -} -.icons-battery-full { - background-position: -224px -288px; -} -.icons-battery-charge { - background-position: -256px -288px; -} -/*11*/ -.icons-pie { - background-position: 0px -320px; -} -.icons-histogram { - background-position: -32px -320px; -} -.icons-window { - background-position: -64px -320px; -} -.icons-window-line{ - background-position: -96px -320px; -} -.icons-window-command{ - background-position: -128px -320px; -} -.icons-window-list{ - background-position: -160px -320px; -} -.icons-window-block{ - background-position: -192px -320px; -} -.icons-terminal{ - background-position: -224px -320px; -} -/*12*/ -.icons-heart-w { - background-position: 0px -352px; -} -.icons-heart-b { - background-position: -32px -352px; -} -.icons-like { - background-position: -64px -352px; -} -.icons-hate { - background-position: -96px -352px; -} -.icons-medal { - background-position: -128px -352px; -} -.icons-warning { - background-position: -160px -352px; -} -.icons-check { - background-position: -192px -352px; -} -.icons-check-box-solid { - background-position: -224px -352px; -} -.icons-check-box-dot { - background-position: -256px -352px; -} -.icons-check-2 { - background-position: -288px -352px; -} -.icons-check-circle-solid { - background-position: -320px -352px; -} -.icons-check-circle-dot { - background-position: -352px -352px; -} -.icons-check-circle-b { - background-position: -384px -352px; -} -.icons-star-thin { - background-position: -416px -352px; -} -.icons-star { - background-position: -448px -352px; -} -/*13*/ -.icons-13-1 { - background-position: -0px -384px; -} -.icons-13-2 { - background-position: -32px -384px; -} -.icons-13-3 { - background-position: -64px -384px; -} -.icons-13-4 { - background-position: -96px -384px; -} -.icons-13-5 { - background-position: -128px -384px; -} -.icons-13-6 { - background-position: -160px -384px; -} -.icons-13-7 { - background-position: -192px -384px; -} -.icons-13-8 { - background-position: -224px -384px; -} -.icons-13-9 { - background-position: -256px -384px; -} -/*14*/ -.icons-14-1 { - background-position: -0px -416px; -} -.icons-14-2 { - background-position: -32px -416px; -} -.icons-14-3 { - background-position: -64px -416px; -} -.icons-14-4 { - background-position: -96px -416px; -} -.icons-14-5 { - background-position: -128px -416px; -} -.icons-14-6 { - background-position: -160px -416px; -} -.icons-14-7 { - background-position: -192px -416px; -} -.icons-14-8 { - background-position: -224px -416px; -} -.icons-14-9 { - background-position: -256px -416px; -} -.icons-14-10 { - background-position: -288px -416px; -} -.icons-14-11 { - background-position: -320px -416px; -} -.icons-14-12 { - background-position: -352px -416px; -} -.icons-14-13 { - background-position: -384px -416px; -} -/*15*/ -.icons-15-1 { - background-position: -0px -448px; -} -.icons-15-2 { - background-position: -32px -448px; -} -.icons-15-3 { - background-position: -64px -448px; -} -.icons-15-4 { - background-position: -96px -448px; -} -.icons-15-5 { - background-position: -128px -448px; -} -.icons-15-6 { - background-position: -160px -448px; -} -.icons-15-7 { - background-position: -192px -448px; -} -.icons-15-8 { - background-position: -224px -448px; -} -.icons-15-9 { - background-position: -256px -448px; -} -.icons-15-10 { - background-position: -288px -448px; -} -.icons-15-11 { - background-position: -320px -448px; -} -.icons-15-12 { - background-position: -352px -448px; -} -.icons-15-13 { - background-position: -384px -448px; -} -.icons-15-14 { - background-position: -416px -448px; -} -.icons-15-15 { - background-position: -448px -448px; -} -.icons-15-16 { - background-position: -480px -448px; -} -.icons-15-17 { - background-position: -512px -448px; -} -/*16*/ -.icons-16-1 { - background-position: -0px -480px; -} -.icons-16-2 { - background-position: -32px -480px; -} -.icons-16-3 { - background-position: -64px -480px; -} -.icons-16-4 { - background-position: -96px -480px; -} -.icons-16-5 { - background-position: -128px -480px; -} -.icons-16-6 { - background-position: -160px -480px; -} -.icons-16-7 { - background-position: -192px -480px; -} -.icons-16-8 { - background-position: -224px -480px; -} -.icons-16-9 { - background-position: -256px -480px; -} -.icons-16-10 { - background-position: -288px -480px; -} -.icons-16-11 { - background-position: -320px -480px; -} -.icons-16-12 { - background-position: -352px -480px; -} -.icons-16-13 { - background-position: -384px -480px; -} -.icons-16-14 { - background-position: -416px -480px; -} -.icons-16-15 { - background-position: -448px -480px; -} -.icons-16-16 { - background-position: -480px -480px; -} -.icons-16-17 { - background-position: -512px -480px; -} -.icons-16-18 { - background-position: -548px -480px; -} -.icons-16-19 { - background-position: -580px -480px; -} -.icons-16-20 { - background-position: -612px -480px; -} -.icons-16-21 { - background-position: -640px -480px; -} -.icons-16-22 { - background-position: -672px -480px; -} -/*17*/ -.icons- { - background-position: -0px -512px; -} -/*18*/ -.icons-help { - background-position: -160px -544px; -} -.icons- { - background-position: -0px -544px; -} -/*19*/ -.icons-plus-cube { - background-position: -192px -576px; -} -.icons-plus { - background-position: -288px -576px; -} -/*20*/ -.icons-power { - background-position: -0px -608px; -} -.icons-output { - background-position: -32px -608px; -} -.icons-col-resize { - background-position: -64px -608px; -} -.icons-move { - background-position: -96px -608px; -} -.icons-size-out { - background-position: -128px -608px; -} -.icons-size-in { - background-position: -160px -608px; -} -.icons-slash { - background-position: -192px -608px; -} -.icons-level { - background-position: -224px -608px; -} -.icons-share { - background-position: -256px -608px; -} -.icons-share2 { - background-position: -288px -608px; -} -.icons-re { - background-position: -320px -608px; -} -.icons-insert { - background-position: -352px -608px; -} -.icons-insert2 { - background-position: -384px -608px; -} -.icons-download { - background-position: -416px -608px; -} -.icons-tag-rignt { - background-position: -448px -608px; -} -.icons-tag-top { - background-position: -480px -608px; -} -.icons-tag-bottom { - background-position: -512px -608px; -} -.icons-tag-left { - background-position: -544px -608px; -} -.icons-moves { - background-position: -576px -608px; -} -/*21*/ -.icons- { - background-position: -0px -640px; -} +} \ No newline at end of file diff --git a/app/helpers/application_helper.rb b/app/helpers/application_helper.rb index 4ad52a40..6f6bd13b 100644 --- a/app/helpers/application_helper.rb +++ b/app/helpers/application_helper.rb @@ -183,6 +183,7 @@ module ApplicationHelper unless edit stylesheets << "\n" stylesheets << "\n" + stylesheets << "\n" end stylesheets << "\n" if page.design.css_reset stylesheets << "\n" diff --git a/app/views/admin/ad_images/_form.html.erb b/app/views/admin/ad_images/_form.html.erb index 008421bc..413ccbf4 100644 --- a/app/views/admin/ad_images/_form.html.erb +++ b/app/views/admin/ad_images/_form.html.erb @@ -19,7 +19,7 @@
    -

    <%= t('nccu.date') %>

    +

    <%= t('nccu.date') %>

    @@ -73,7 +73,7 @@
    -

    <%= t('nccu.picture') %>

    +

    <%= t('nccu.picture') %>

    @@ -101,7 +101,7 @@
    -

    <%= t(:type) %>

    +

    <%= t(:type) %>

    <%= f.select :link_open ,AdImage::LINK_OPEN_TYPES%>
    @@ -111,7 +111,7 @@
    -

    <%= t(:frequency) %>

    +

    <%= t(:frequency) %>

    <%= f.text_field :weight ,:class=> 'span3',:placeholder=>"在套圖中出現次數 1次請輸入1" %>
    @@ -121,7 +121,7 @@
    -

    <%= t(:link) %>

    +

    <%= t(:link) %>

    <%= f.text_field :out_link ,:class=> 'span3',:placeholder => "輸入連結"%>
    diff --git a/app/views/admin/assets/_asset.html.erb b/app/views/admin/assets/_asset.html.erb index 3040f872..c67ade0e 100644 --- a/app/views/admin/assets/_asset.html.erb +++ b/app/views/admin/assets/_asset.html.erb @@ -2,7 +2,7 @@ <%= check_box_tag 'to_delete[]', asset.id, false, :class => "checkbox_in_list" %> <%= asset.title rescue nil %> - +